xtreme

References:

XNLI

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/XNLI')
  • Description:
The Cross-lingual Natural Language Inference (XNLI) corpus is a crowd-sourced collection of 5,000 test and
2,500 dev pairs for the MultiNLI corpus. The pairs are annotated with textual entailment and translated into
14 languages: French, Spanish, German, Greek, Bulgarian, Russian, Turkish, Arabic, Vietnamese, Thai, Chinese,
Hindi, Swahili and Urdu. This results in 112.5k annotated pairs. Each premise can be associated with the
corresponding hypothesis in the 15 languages, summing up to more than 1.5M combinations. The corpus is made to
evaluate how to perform inference in any language (including low-resources ones like Swahili or Urdu) when only
English NLI data is available at training time. One solution is cross-lingual sentence encoding, for which XNLI
is an evaluation benchmark.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 75150
'validation' 37350
  • Features:
{
    "language": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "sentence1": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "sentence2": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "gold_label": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    }
}

tydiqa

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/tydiqa')
  • Description:
Gold passage task (GoldP): Given a passage that is guaranteed to contain the
             answer, predict the single contiguous span of characters that answers the question. This is more similar to
             existing reading comprehension datasets (as opposed to the information-seeking task outlined above).
             This task is constructed with two goals in mind: (1) more directly comparing with prior work and (2) providing
             a simplified way for researchers to use TyDi QA by providing compatibility with existing code for SQuAD 1.1,
             XQuAD, and MLQA. Toward these goals, the gold passage task differs from the primary task in several ways:
             only the gold answer passage is provided rather than the entire Wikipedia article;
             unanswerable questions have been discarded, similar to MLQA and XQuAD;
             we evaluate with the SQuAD 1.1 metrics like XQuAD; and
            Thai and Japanese are removed since the lack of whitespace breaks some tools.

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'train' 49881
'validation' 5077
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "title": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

SQuAD

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/SQuAD')
  • Description:
Stanford Question Answering Dataset (SQuAD) is a reading comprehension     dataset, consisting of questions posed by crowdworkers on a set of Wikipedia     articles, where the answer to every question is a segment of text, or span,     from the corresponding reading passage, or the question might be unanswerable.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'train' 87599
'validation' 10570
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "title": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

PAN-X.af

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/PAN-X.af')
  • Description:
The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been
constructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset
can be loaded with the DaNLP package:
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 1000
'train' 5000
'validation' 1000
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "ner_tags": {
        "feature": {
            "num_classes": 7,
            "names": [
                "O",
                "B-PER",
                "I-PER",
                "B-ORG",
                "I-ORG",
                "B-LOC",
                "I-LOC"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "langs": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

PAN-X.ar

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/PAN-X.ar')
  • Description:
The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been
constructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset
can be loaded with the DaNLP package:
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 10000
'train' 20000
'validation' 10000
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "ner_tags": {
        "feature": {
            "num_classes": 7,
            "names": [
                "O",
                "B-PER",
                "I-PER",
                "B-ORG",
                "I-ORG",
                "B-LOC",
                "I-LOC"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "langs": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

PAN-X.bg

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/PAN-X.bg')
  • Description:
The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been
constructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset
can be loaded with the DaNLP package:
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 10000
'train' 20000
'validation' 10000
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "ner_tags": {
        "feature": {
            "num_classes": 7,
            "names": [
                "O",
                "B-PER",
                "I-PER",
                "B-ORG",
                "I-ORG",
                "B-LOC",
                "I-LOC"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "langs": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

PAN-X.bn

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/PAN-X.bn')
  • Description:
The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been
constructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset
can be loaded with the DaNLP package:
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 1000
'train' 10000
'validation' 1000
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "ner_tags": {
        "feature": {
            "num_classes": 7,
            "names": [
                "O",
                "B-PER",
                "I-PER",
                "B-ORG",
                "I-ORG",
                "B-LOC",
                "I-LOC"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "langs": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

PAN-X.de

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/PAN-X.de')
  • Description:
The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been
constructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset
can be loaded with the DaNLP package:
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 10000
'train' 20000
'validation' 10000
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "ner_tags": {
        "feature": {
            "num_classes": 7,
            "names": [
                "O",
                "B-PER",
                "I-PER",
                "B-ORG",
                "I-ORG",
                "B-LOC",
                "I-LOC"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "langs": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

PAN-X.el

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/PAN-X.el')
  • Description:
The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been
constructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset
can be loaded with the DaNLP package:
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 10000
'train' 20000
'validation' 10000
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "ner_tags": {
        "feature": {
            "num_classes": 7,
            "names": [
                "O",
                "B-PER",
                "I-PER",
                "B-ORG",
                "I-ORG",
                "B-LOC",
                "I-LOC"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "langs": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

PAN-X.en

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/PAN-X.en')
  • Description:
The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been
constructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset
can be loaded with the DaNLP package:
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 10000
'train' 20000
'validation' 10000
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "ner_tags": {
        "feature": {
            "num_classes": 7,
            "names": [
                "O",
                "B-PER",
                "I-PER",
                "B-ORG",
                "I-ORG",
                "B-LOC",
                "I-LOC"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "langs": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

PAN-X.es

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/PAN-X.es')
  • Description:
The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been
constructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset
can be loaded with the DaNLP package:
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 10000
'train' 20000
'validation' 10000
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "ner_tags": {
        "feature": {
            "num_classes": 7,
            "names": [
                "O",
                "B-PER",
                "I-PER",
                "B-ORG",
                "I-ORG",
                "B-LOC",
                "I-LOC"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "langs": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

PAN-X.et

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/PAN-X.et')
  • Description:
The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been
constructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset
can be loaded with the DaNLP package:
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 10000
'train' 15000
'validation' 10000
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "ner_tags": {
        "feature": {
            "num_classes": 7,
            "names": [
                "O",
                "B-PER",
                "I-PER",
                "B-ORG",
                "I-ORG",
                "B-LOC",
                "I-LOC"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "langs": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

PAN-X.eu

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/PAN-X.eu')
  • Description:
The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been
constructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset
can be loaded with the DaNLP package:
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 10000
'train' 10000
'validation' 10000
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "ner_tags": {
        "feature": {
            "num_classes": 7,
            "names": [
                "O",
                "B-PER",
                "I-PER",
                "B-ORG",
                "I-ORG",
                "B-LOC",
                "I-LOC"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "langs": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

PAN-X.fa

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/PAN-X.fa')
  • Description:
The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been
constructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset
can be loaded with the DaNLP package:
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 10000
'train' 20000
'validation' 10000
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "ner_tags": {
        "feature": {
            "num_classes": 7,
            "names": [
                "O",
                "B-PER",
                "I-PER",
                "B-ORG",
                "I-ORG",
                "B-LOC",
                "I-LOC"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "langs": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

PAN-X.fi

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/PAN-X.fi')
  • Description:
The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been
constructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset
can be loaded with the DaNLP package:
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 10000
'train' 20000
'validation' 10000
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "ner_tags": {
        "feature": {
            "num_classes": 7,
            "names": [
                "O",
                "B-PER",
                "I-PER",
                "B-ORG",
                "I-ORG",
                "B-LOC",
                "I-LOC"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "langs": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

PAN-X.fr

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/PAN-X.fr')
  • Description:
The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been
constructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset
can be loaded with the DaNLP package:
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 10000
'train' 20000
'validation' 10000
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "ner_tags": {
        "feature": {
            "num_classes": 7,
            "names": [
                "O",
                "B-PER",
                "I-PER",
                "B-ORG",
                "I-ORG",
                "B-LOC",
                "I-LOC"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "langs": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

PAN-X.he

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/PAN-X.he')
  • Description:
The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been
constructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset
can be loaded with the DaNLP package:
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 10000
'train' 20000
'validation' 10000
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "ner_tags": {
        "feature": {
            "num_classes": 7,
            "names": [
                "O",
                "B-PER",
                "I-PER",
                "B-ORG",
                "I-ORG",
                "B-LOC",
                "I-LOC"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "langs": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

PAN-X.hi

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/PAN-X.hi')
  • Description:
The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been
constructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset
can be loaded with the DaNLP package:
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 1000
'train' 5000
'validation' 1000
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "ner_tags": {
        "feature": {
            "num_classes": 7,
            "names": [
                "O",
                "B-PER",
                "I-PER",
                "B-ORG",
                "I-ORG",
                "B-LOC",
                "I-LOC"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "langs": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

PAN-X.hu

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/PAN-X.hu')
  • Description:
The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been
constructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset
can be loaded with the DaNLP package:
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 10000
'train' 20000
'validation' 10000
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "ner_tags": {
        "feature": {
            "num_classes": 7,
            "names": [
                "O",
                "B-PER",
                "I-PER",
                "B-ORG",
                "I-ORG",
                "B-LOC",
                "I-LOC"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "langs": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

PAN-X.id

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/PAN-X.id')
  • Description:
The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been
constructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset
can be loaded with the DaNLP package:
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 10000
'train' 20000
'validation' 10000
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "ner_tags": {
        "feature": {
            "num_classes": 7,
            "names": [
                "O",
                "B-PER",
                "I-PER",
                "B-ORG",
                "I-ORG",
                "B-LOC",
                "I-LOC"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "langs": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

PAN-X.it

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/PAN-X.it')
  • Description:
The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been
constructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset
can be loaded with the DaNLP package:
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 10000
'train' 20000
'validation' 10000
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "ner_tags": {
        "feature": {
            "num_classes": 7,
            "names": [
                "O",
                "B-PER",
                "I-PER",
                "B-ORG",
                "I-ORG",
                "B-LOC",
                "I-LOC"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "langs": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

PAN-X.ja

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/PAN-X.ja')
  • Description:
The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been
constructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset
can be loaded with the DaNLP package:
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 10000
'train' 20000
'validation' 10000
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "ner_tags": {
        "feature": {
            "num_classes": 7,
            "names": [
                "O",
                "B-PER",
                "I-PER",
                "B-ORG",
                "I-ORG",
                "B-LOC",
                "I-LOC"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "langs": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

PAN-X.jv

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/PAN-X.jv')
  • Description:
The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been
constructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset
can be loaded with the DaNLP package:
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 100
'train' 100
'validation' 100
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "ner_tags": {
        "feature": {
            "num_classes": 7,
            "names": [
                "O",
                "B-PER",
                "I-PER",
                "B-ORG",
                "I-ORG",
                "B-LOC",
                "I-LOC"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "langs": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

PAN-X.ka

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/PAN-X.ka')
  • Description:
The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been
constructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset
can be loaded with the DaNLP package:
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 10000
'train' 10000
'validation' 10000
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "ner_tags": {
        "feature": {
            "num_classes": 7,
            "names": [
                "O",
                "B-PER",
                "I-PER",
                "B-ORG",
                "I-ORG",
                "B-LOC",
                "I-LOC"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "langs": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

PAN-X.kk

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/PAN-X.kk')
  • Description:
The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been
constructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset
can be loaded with the DaNLP package:
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 1000
'train' 1000
'validation' 1000
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "ner_tags": {
        "feature": {
            "num_classes": 7,
            "names": [
                "O",
                "B-PER",
                "I-PER",
                "B-ORG",
                "I-ORG",
                "B-LOC",
                "I-LOC"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "langs": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

PAN-X.ko

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/PAN-X.ko')
  • Description:
The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been
constructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset
can be loaded with the DaNLP package:
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 10000
'train' 20000
'validation' 10000
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "ner_tags": {
        "feature": {
            "num_classes": 7,
            "names": [
                "O",
                "B-PER",
                "I-PER",
                "B-ORG",
                "I-ORG",
                "B-LOC",
                "I-LOC"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "langs": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

PAN-X.ml

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/PAN-X.ml')
  • Description:
The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been
constructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset
can be loaded with the DaNLP package:
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 1000
'train' 10000
'validation' 1000
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "ner_tags": {
        "feature": {
            "num_classes": 7,
            "names": [
                "O",
                "B-PER",
                "I-PER",
                "B-ORG",
                "I-ORG",
                "B-LOC",
                "I-LOC"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "langs": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

PAN-X.mr

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/PAN-X.mr')
  • Description:
The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been
constructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset
can be loaded with the DaNLP package:
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 1000
'train' 5000
'validation' 1000
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "ner_tags": {
        "feature": {
            "num_classes": 7,
            "names": [
                "O",
                "B-PER",
                "I-PER",
                "B-ORG",
                "I-ORG",
                "B-LOC",
                "I-LOC"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "langs": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

PAN-X.ms

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/PAN-X.ms')
  • Description:
The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been
constructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset
can be loaded with the DaNLP package:
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 1000
'train' 20000
'validation' 1000
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "ner_tags": {
        "feature": {
            "num_classes": 7,
            "names": [
                "O",
                "B-PER",
                "I-PER",
                "B-ORG",
                "I-ORG",
                "B-LOC",
                "I-LOC"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "langs": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

PAN-X.my

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/PAN-X.my')
  • Description:
The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been
constructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset
can be loaded with the DaNLP package:
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 100
'train' 100
'validation' 100
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "ner_tags": {
        "feature": {
            "num_classes": 7,
            "names": [
                "O",
                "B-PER",
                "I-PER",
                "B-ORG",
                "I-ORG",
                "B-LOC",
                "I-LOC"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "langs": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

PAN-X.nl

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/PAN-X.nl')
  • Description:
The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been
constructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset
can be loaded with the DaNLP package:
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 10000
'train' 20000
'validation' 10000
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "ner_tags": {
        "feature": {
            "num_classes": 7,
            "names": [
                "O",
                "B-PER",
                "I-PER",
                "B-ORG",
                "I-ORG",
                "B-LOC",
                "I-LOC"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "langs": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

PAN-X.pt

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/PAN-X.pt')
  • Description:
The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been
constructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset
can be loaded with the DaNLP package:
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 10000
'train' 20000
'validation' 10000
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "ner_tags": {
        "feature": {
            "num_classes": 7,
            "names": [
                "O",
                "B-PER",
                "I-PER",
                "B-ORG",
                "I-ORG",
                "B-LOC",
                "I-LOC"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "langs": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

PAN-X.ru

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/PAN-X.ru')
  • Description:
The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been
constructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset
can be loaded with the DaNLP package:
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 10000
'train' 20000
'validation' 10000
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "ner_tags": {
        "feature": {
            "num_classes": 7,
            "names": [
                "O",
                "B-PER",
                "I-PER",
                "B-ORG",
                "I-ORG",
                "B-LOC",
                "I-LOC"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "langs": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

PAN-X.sw

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/PAN-X.sw')
  • Description:
The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been
constructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset
can be loaded with the DaNLP package:
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 1000
'train' 1000
'validation' 1000
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "ner_tags": {
        "feature": {
            "num_classes": 7,
            "names": [
                "O",
                "B-PER",
                "I-PER",
                "B-ORG",
                "I-ORG",
                "B-LOC",
                "I-LOC"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "langs": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

PAN-X.ta

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/PAN-X.ta')
  • Description:
The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been
constructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset
can be loaded with the DaNLP package:
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 1000
'train' 15000
'validation' 1000
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "ner_tags": {
        "feature": {
            "num_classes": 7,
            "names": [
                "O",
                "B-PER",
                "I-PER",
                "B-ORG",
                "I-ORG",
                "B-LOC",
                "I-LOC"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "langs": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

PAN-X.te

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/PAN-X.te')
  • Description:
The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been
constructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset
can be loaded with the DaNLP package:
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 1000
'train' 1000
'validation' 1000
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "ner_tags": {
        "feature": {
            "num_classes": 7,
            "names": [
                "O",
                "B-PER",
                "I-PER",
                "B-ORG",
                "I-ORG",
                "B-LOC",
                "I-LOC"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "langs": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

PAN-X.th

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/PAN-X.th')
  • Description:
The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been
constructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset
can be loaded with the DaNLP package:
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 10000
'train' 20000
'validation' 10000
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "ner_tags": {
        "feature": {
            "num_classes": 7,
            "names": [
                "O",
                "B-PER",
                "I-PER",
                "B-ORG",
                "I-ORG",
                "B-LOC",
                "I-LOC"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "langs": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

PAN-X.tl

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/PAN-X.tl')
  • Description:
The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been
constructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset
can be loaded with the DaNLP package:
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 1000
'train' 10000
'validation' 1000
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "ner_tags": {
        "feature": {
            "num_classes": 7,
            "names": [
                "O",
                "B-PER",
                "I-PER",
                "B-ORG",
                "I-ORG",
                "B-LOC",
                "I-LOC"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "langs": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

PAN-X.tr

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/PAN-X.tr')
  • Description:
The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been
constructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset
can be loaded with the DaNLP package:
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 10000
'train' 20000
'validation' 10000
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "ner_tags": {
        "feature": {
            "num_classes": 7,
            "names": [
                "O",
                "B-PER",
                "I-PER",
                "B-ORG",
                "I-ORG",
                "B-LOC",
                "I-LOC"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "langs": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

PAN-X.ur

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/PAN-X.ur')
  • Description:
The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been
constructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset
can be loaded with the DaNLP package:
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 1000
'train' 20000
'validation' 1000
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "ner_tags": {
        "feature": {
            "num_classes": 7,
            "names": [
                "O",
                "B-PER",
                "I-PER",
                "B-ORG",
                "I-ORG",
                "B-LOC",
                "I-LOC"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "langs": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

PAN-X.vi

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/PAN-X.vi')
  • Description:
The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been
constructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset
can be loaded with the DaNLP package:
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 10000
'train' 20000
'validation' 10000
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "ner_tags": {
        "feature": {
            "num_classes": 7,
            "names": [
                "O",
                "B-PER",
                "I-PER",
                "B-ORG",
                "I-ORG",
                "B-LOC",
                "I-LOC"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "langs": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

PAN-X.yo

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/PAN-X.yo')
  • Description:
The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been
constructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset
can be loaded with the DaNLP package:
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 100
'train' 100
'validation' 100
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "ner_tags": {
        "feature": {
            "num_classes": 7,
            "names": [
                "O",
                "B-PER",
                "I-PER",
                "B-ORG",
                "I-ORG",
                "B-LOC",
                "I-LOC"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "langs": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

PAN-X.zh

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/PAN-X.zh')
  • Description:
The WikiANN dataset (Pan et al. 2017) is a dataset with NER annotations for PER, ORG and LOC. It has been
constructed using the linked entities in Wikipedia pages for 282 different languages including Danish. The dataset
can be loaded with the DaNLP package:
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 10000
'train' 20000
'validation' 10000
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "ner_tags": {
        "feature": {
            "num_classes": 7,
            "names": [
                "O",
                "B-PER",
                "I-PER",
                "B-ORG",
                "I-ORG",
                "B-LOC",
                "I-LOC"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "langs": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

MLQA.ar.ar

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/MLQA.ar.ar')
  • Description:
MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.
MLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,
German, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between
4 different languages on average.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 5335
'validation' 517
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "title": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

MLQA.ar.de

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/MLQA.ar.de')
  • Description:
MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.
MLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,
German, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between
4 different languages on average.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 1649
'validation' 207
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "title": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

MLQA.ar.vi

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/MLQA.ar.vi')
  • Description:
MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.
MLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,
German, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between
4 different languages on average.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 2047
'validation' 163
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "title": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

MLQA.ar.zh

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/MLQA.ar.zh')
  • Description:
MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.
MLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,
German, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between
4 different languages on average.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 1912
'validation' 188
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "title": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

MLQA.ar.en

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/MLQA.ar.en')
  • Description:
MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.
MLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,
German, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between
4 different languages on average.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 5335
'validation' 517
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "title": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

MLQA.ar.es

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/MLQA.ar.es')
  • Description:
MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.
MLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,
German, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between
4 different languages on average.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 1978
'validation' 161
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "title": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

MLQA.ar.hi

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/MLQA.ar.hi')
  • Description:
MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.
MLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,
German, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between
4 different languages on average.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 1831
'validation' 186
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "title": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

MLQA.de.ar

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/MLQA.de.ar')
  • Description:
MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.
MLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,
German, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between
4 different languages on average.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 1649
'validation' 207
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "title": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

MLQA.de.de

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/MLQA.de.de')
  • Description:
MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.
MLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,
German, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between
4 different languages on average.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 4517
'validation' 512
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "title": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

MLQA.de.vi

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/MLQA.de.vi')
  • Description:
MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.
MLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,
German, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between
4 different languages on average.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 1675
'validation' 182
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "title": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

MLQA.de.zh

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/MLQA.de.zh')
  • Description:
MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.
MLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,
German, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between
4 different languages on average.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 1621
'validation' 190
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "title": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

MLQA.de.en

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/MLQA.de.en')
  • Description:
MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.
MLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,
German, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between
4 different languages on average.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 4517
'validation' 512
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "title": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

MLQA.de.es

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/MLQA.de.es')
  • Description:
MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.
MLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,
German, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between
4 different languages on average.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 1776
'validation' 196
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "title": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

MLQA.de.hi

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/MLQA.de.hi')
  • Description:
MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.
MLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,
German, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between
4 different languages on average.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 1430
'validation' 163
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "title": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

MLQA.vi.ar

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/MLQA.vi.ar')
  • Description:
MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.
MLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,
German, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between
4 different languages on average.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 2047
'validation' 163
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "title": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

MLQA.vi.de

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/MLQA.vi.de')
  • Description:
MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.
MLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,
German, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between
4 different languages on average.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 1675
'validation' 182
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "title": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

MLQA.vi.vi

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/MLQA.vi.vi')
  • Description:
MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.
MLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,
German, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between
4 different languages on average.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 5495
'validation' 511
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "title": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

MLQA.vi.zh

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/MLQA.vi.zh')
  • Description:
MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.
MLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,
German, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between
4 different languages on average.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 1943
'validation' 184
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "title": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

MLQA.vi.en

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/MLQA.vi.en')
  • Description:
MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.
MLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,
German, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between
4 different languages on average.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 5495
'validation' 511
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "title": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

MLQA.vi.es

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/MLQA.vi.es')
  • Description:
MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.
MLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,
German, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between
4 different languages on average.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 2018
'validation' 189
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "title": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

MLQA.vi.hi

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/MLQA.vi.hi')
  • Description:
MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.
MLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,
German, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between
4 different languages on average.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 1947
'validation' 177
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "title": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

MLQA.zh.ar

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/MLQA.zh.ar')
  • Description:
MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.
MLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,
German, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between
4 different languages on average.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 1912
'validation' 188
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "title": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

MLQA.zh.de

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/MLQA.zh.de')
  • Description:
MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.
MLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,
German, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between
4 different languages on average.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 1621
'validation' 190
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "title": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

MLQA.zh.vi

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/MLQA.zh.vi')
  • Description:
MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.
MLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,
German, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between
4 different languages on average.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 1943
'validation' 184
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "title": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

MLQA.zh.zh

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/MLQA.zh.zh')
  • Description:
MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.
MLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,
German, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between
4 different languages on average.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 5137
'validation' 504
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "title": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

MLQA.zh.en

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/MLQA.zh.en')
  • Description:
MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.
MLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,
German, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between
4 different languages on average.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 5137
'validation' 504
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "title": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

MLQA.zh.es

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/MLQA.zh.es')
  • Description:
MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.
MLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,
German, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between
4 different languages on average.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 1947
'validation' 161
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "title": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

MLQA.zh.hi

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/MLQA.zh.hi')
  • Description:
MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.
MLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,
German, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between
4 different languages on average.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 1767
'validation' 189
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "title": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

MLQA.en.ar

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/MLQA.en.ar')
  • Description:
MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.
MLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,
German, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between
4 different languages on average.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 5335
'validation' 517
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "title": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

MLQA.en.de

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/MLQA.en.de')
  • Description:
MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.
MLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,
German, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between
4 different languages on average.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 4517
'validation' 512
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "title": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

MLQA.en.vi

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/MLQA.en.vi')
  • Description:
MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.
MLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,
German, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between
4 different languages on average.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 5495
'validation' 511
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "title": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

MLQA.en.zh

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/MLQA.en.zh')
  • Description:
MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.
MLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,
German, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between
4 different languages on average.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 5137
'validation' 504
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "title": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

MLQA.en.en

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/MLQA.en.en')
  • Description:
MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.
MLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,
German, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between
4 different languages on average.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 11590
'validation' 1148
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "title": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

MLQA.en.es

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/MLQA.en.es')
  • Description:
MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.
MLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,
German, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between
4 different languages on average.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 5253
'validation' 500
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "title": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

MLQA.en.hi

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/MLQA.en.hi')
  • Description:
MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.
MLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,
German, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between
4 different languages on average.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 4918
'validation' 507
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "title": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

MLQA.es.ar

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/MLQA.es.ar')
  • Description:
MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.
MLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,
German, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between
4 different languages on average.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 1978
'validation' 161
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "title": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

MLQA.es.de

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/MLQA.es.de')
  • Description:
MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.
MLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,
German, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between
4 different languages on average.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 1776
'validation' 196
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "title": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

MLQA.es.vi

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/MLQA.es.vi')
  • Description:
MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.
MLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,
German, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between
4 different languages on average.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 2018
'validation' 189
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "title": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

MLQA.es.zh

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/MLQA.es.zh')
  • Description:
MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.
MLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,
German, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between
4 different languages on average.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 1947
'validation' 161
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "title": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

MLQA.es.en

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/MLQA.es.en')
  • Description:
MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.
MLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,
German, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between
4 different languages on average.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 5253
'validation' 500
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "title": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

MLQA.es.es

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/MLQA.es.es')
  • Description:
MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.
MLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,
German, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between
4 different languages on average.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 5253
'validation' 500
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "title": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

MLQA.es.hi

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/MLQA.es.hi')
  • Description:
MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.
MLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,
German, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between
4 different languages on average.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 1723
'validation' 187
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "title": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

MLQA.hi.ar

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/MLQA.hi.ar')
  • Description:
MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.
MLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,
German, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between
4 different languages on average.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 1831
'validation' 186
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "title": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

MLQA.hi.de

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/MLQA.hi.de')
  • Description:
MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.
MLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,
German, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between
4 different languages on average.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 1430
'validation' 163
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "title": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

MLQA.hi.vi

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/MLQA.hi.vi')
  • Description:
MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.
MLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,
German, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between
4 different languages on average.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 1947
'validation' 177
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "title": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

MLQA.hi.zh

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/MLQA.hi.zh')
  • Description:
MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.
MLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,
German, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between
4 different languages on average.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 1767
'validation' 189
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "title": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

MLQA.hi.en

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/MLQA.hi.en')
  • Description:
MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.
MLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,
German, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between
4 different languages on average.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 4918
'validation' 507
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "title": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

MLQA.hi.es

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/MLQA.hi.es')
  • Description:
MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.
MLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,
German, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between
4 different languages on average.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 1723
'validation' 187
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "title": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

MLQA.hi.hi

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/MLQA.hi.hi')
  • Description:
MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.
MLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,
German, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between
4 different languages on average.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 4918
'validation' 507
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "title": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

XQuAD.ar

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/XQuAD.ar')
  • Description:
XQuAD (Cross-lingual Question Answering Dataset) is a benchmark dataset for evaluating cross-lingual question
answering performance. The dataset consists of a subset of 240 paragraphs and 1190 question-answer pairs from
the development set of SQuAD v1.1 (Rajpurkar et al., 2016) together with their professional translations into
ten languages: Spanish, German, Greek, Russian, Turkish, Arabic, Vietnamese, Thai, Chinese, and Hindi. Consequently,
the dataset is entirely parallel across 11 languages.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'validation' 1190
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

XQuAD.de

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/XQuAD.de')
  • Description:
XQuAD (Cross-lingual Question Answering Dataset) is a benchmark dataset for evaluating cross-lingual question
answering performance. The dataset consists of a subset of 240 paragraphs and 1190 question-answer pairs from
the development set of SQuAD v1.1 (Rajpurkar et al., 2016) together with their professional translations into
ten languages: Spanish, German, Greek, Russian, Turkish, Arabic, Vietnamese, Thai, Chinese, and Hindi. Consequently,
the dataset is entirely parallel across 11 languages.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'validation' 1190
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

XQuAD.vi

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/XQuAD.vi')
  • Description:
XQuAD (Cross-lingual Question Answering Dataset) is a benchmark dataset for evaluating cross-lingual question
answering performance. The dataset consists of a subset of 240 paragraphs and 1190 question-answer pairs from
the development set of SQuAD v1.1 (Rajpurkar et al., 2016) together with their professional translations into
ten languages: Spanish, German, Greek, Russian, Turkish, Arabic, Vietnamese, Thai, Chinese, and Hindi. Consequently,
the dataset is entirely parallel across 11 languages.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'validation' 1190
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

XQuAD.zh

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/XQuAD.zh')
  • Description:
XQuAD (Cross-lingual Question Answering Dataset) is a benchmark dataset for evaluating cross-lingual question
answering performance. The dataset consists of a subset of 240 paragraphs and 1190 question-answer pairs from
the development set of SQuAD v1.1 (Rajpurkar et al., 2016) together with their professional translations into
ten languages: Spanish, German, Greek, Russian, Turkish, Arabic, Vietnamese, Thai, Chinese, and Hindi. Consequently,
the dataset is entirely parallel across 11 languages.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'validation' 1190
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

XQuAD.en

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/XQuAD.en')
  • Description:
XQuAD (Cross-lingual Question Answering Dataset) is a benchmark dataset for evaluating cross-lingual question
answering performance. The dataset consists of a subset of 240 paragraphs and 1190 question-answer pairs from
the development set of SQuAD v1.1 (Rajpurkar et al., 2016) together with their professional translations into
ten languages: Spanish, German, Greek, Russian, Turkish, Arabic, Vietnamese, Thai, Chinese, and Hindi. Consequently,
the dataset is entirely parallel across 11 languages.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'validation' 1190
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

XQuAD.es

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/XQuAD.es')
  • Description:
XQuAD (Cross-lingual Question Answering Dataset) is a benchmark dataset for evaluating cross-lingual question
answering performance. The dataset consists of a subset of 240 paragraphs and 1190 question-answer pairs from
the development set of SQuAD v1.1 (Rajpurkar et al., 2016) together with their professional translations into
ten languages: Spanish, German, Greek, Russian, Turkish, Arabic, Vietnamese, Thai, Chinese, and Hindi. Consequently,
the dataset is entirely parallel across 11 languages.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'validation' 1190
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

XQuAD.hi

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/XQuAD.hi')
  • Description:
XQuAD (Cross-lingual Question Answering Dataset) is a benchmark dataset for evaluating cross-lingual question
answering performance. The dataset consists of a subset of 240 paragraphs and 1190 question-answer pairs from
the development set of SQuAD v1.1 (Rajpurkar et al., 2016) together with their professional translations into
ten languages: Spanish, German, Greek, Russian, Turkish, Arabic, Vietnamese, Thai, Chinese, and Hindi. Consequently,
the dataset is entirely parallel across 11 languages.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'validation' 1190
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

XQuAD.el

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/XQuAD.el')
  • Description:
XQuAD (Cross-lingual Question Answering Dataset) is a benchmark dataset for evaluating cross-lingual question
answering performance. The dataset consists of a subset of 240 paragraphs and 1190 question-answer pairs from
the development set of SQuAD v1.1 (Rajpurkar et al., 2016) together with their professional translations into
ten languages: Spanish, German, Greek, Russian, Turkish, Arabic, Vietnamese, Thai, Chinese, and Hindi. Consequently,
the dataset is entirely parallel across 11 languages.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'validation' 1190
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

XQuAD.ru

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/XQuAD.ru')
  • Description:
XQuAD (Cross-lingual Question Answering Dataset) is a benchmark dataset for evaluating cross-lingual question
answering performance. The dataset consists of a subset of 240 paragraphs and 1190 question-answer pairs from
the development set of SQuAD v1.1 (Rajpurkar et al., 2016) together with their professional translations into
ten languages: Spanish, German, Greek, Russian, Turkish, Arabic, Vietnamese, Thai, Chinese, and Hindi. Consequently,
the dataset is entirely parallel across 11 languages.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'validation' 1190
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

XQuAD.th

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/XQuAD.th')
  • Description:
XQuAD (Cross-lingual Question Answering Dataset) is a benchmark dataset for evaluating cross-lingual question
answering performance. The dataset consists of a subset of 240 paragraphs and 1190 question-answer pairs from
the development set of SQuAD v1.1 (Rajpurkar et al., 2016) together with their professional translations into
ten languages: Spanish, German, Greek, Russian, Turkish, Arabic, Vietnamese, Thai, Chinese, and Hindi. Consequently,
the dataset is entirely parallel across 11 languages.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'validation' 1190
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

XQuAD.tr

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/XQuAD.tr')
  • Description:
XQuAD (Cross-lingual Question Answering Dataset) is a benchmark dataset for evaluating cross-lingual question
answering performance. The dataset consists of a subset of 240 paragraphs and 1190 question-answer pairs from
the development set of SQuAD v1.1 (Rajpurkar et al., 2016) together with their professional translations into
ten languages: Spanish, German, Greek, Russian, Turkish, Arabic, Vietnamese, Thai, Chinese, and Hindi. Consequently,
the dataset is entirely parallel across 11 languages.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'validation' 1190
  • Features:
{
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "context": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "question": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "answers": {
        "feature": {
            "answer_start": {
                "dtype": "int32",
                "id": null,
                "_type": "Value"
            },
            "text": {
                "dtype": "string",
                "id": null,
                "_type": "Value"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

bucc18.de

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/bucc18.de')
  • Description:
Building and Using Comparable Corpora

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 9580
'validation' 1038
  • Features:
{
    "source_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "source_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    }
}

bucc18.fr

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/bucc18.fr')
  • Description:
Building and Using Comparable Corpora

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 9086
'validation' 929
  • Features:
{
    "source_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "source_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    }
}

bucc18.zh

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/bucc18.zh')
  • Description:
Building and Using Comparable Corpora

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 1899
'validation' 257
  • Features:
{
    "source_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "source_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    }
}

bucc18.ru

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/bucc18.ru')
  • Description:
Building and Using Comparable Corpora

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 14435
'validation' 2374
  • Features:
{
    "source_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "source_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    }
}

PAWS-X.de

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/PAWS-X.de')
  • Description:
This dataset contains 23,659 human translated PAWS evaluation pairs and 296,406 machine translated training
pairs in six typologically distinct languages: French, Spanish, German, Chinese, Japanese, and Korean. All
translated pairs are sourced from examples in PAWS-Wiki.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 2000
'train' 49380
'validation' 2000
  • Features:
{
    "sentence1": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "sentence2": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "label": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    }
}

PAWS-X.en

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/PAWS-X.en')
  • Description:
This dataset contains 23,659 human translated PAWS evaluation pairs and 296,406 machine translated training
pairs in six typologically distinct languages: French, Spanish, German, Chinese, Japanese, and Korean. All
translated pairs are sourced from examples in PAWS-Wiki.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 2000
'train' 49175
'validation' 2000
  • Features:
{
    "sentence1": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "sentence2": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "label": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    }
}

PAWS-X.es

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/PAWS-X.es')
  • Description:
This dataset contains 23,659 human translated PAWS evaluation pairs and 296,406 machine translated training
pairs in six typologically distinct languages: French, Spanish, German, Chinese, Japanese, and Korean. All
translated pairs are sourced from examples in PAWS-Wiki.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 2000
'train' 49401
'validation' 1961
  • Features:
{
    "sentence1": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "sentence2": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "label": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    }
}

PAWS-X.fr

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/PAWS-X.fr')
  • Description:
This dataset contains 23,659 human translated PAWS evaluation pairs and 296,406 machine translated training
pairs in six typologically distinct languages: French, Spanish, German, Chinese, Japanese, and Korean. All
translated pairs are sourced from examples in PAWS-Wiki.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 2000
'train' 49399
'validation' 1988
  • Features:
{
    "sentence1": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "sentence2": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "label": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    }
}

PAWS-X.ja

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/PAWS-X.ja')
  • Description:
This dataset contains 23,659 human translated PAWS evaluation pairs and 296,406 machine translated training
pairs in six typologically distinct languages: French, Spanish, German, Chinese, Japanese, and Korean. All
translated pairs are sourced from examples in PAWS-Wiki.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 2000
'train' 49401
'validation' 2000
  • Features:
{
    "sentence1": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "sentence2": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "label": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    }
}

PAWS-X.ko

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/PAWS-X.ko')
  • Description:
This dataset contains 23,659 human translated PAWS evaluation pairs and 296,406 machine translated training
pairs in six typologically distinct languages: French, Spanish, German, Chinese, Japanese, and Korean. All
translated pairs are sourced from examples in PAWS-Wiki.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 1999
'train' 49164
'validation' 2000
  • Features:
{
    "sentence1": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "sentence2": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "label": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    }
}

PAWS-X.zh

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/PAWS-X.zh')
  • Description:
This dataset contains 23,659 human translated PAWS evaluation pairs and 296,406 machine translated training
pairs in six typologically distinct languages: French, Spanish, German, Chinese, Japanese, and Korean. All
translated pairs are sourced from examples in PAWS-Wiki.
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 2000
'train' 49401
'validation' 2000
  • Features:
{
    "sentence1": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "sentence2": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "label": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    }
}

tatoeba.afr

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/tatoeba.afr')
  • Description:
his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.
For each languages, we have selected 1000 English sentences and their translations, if available. Please check
this paper for a description of the languages, their families and scripts as well as baseline results.
Please note that the English sentences are not identical for all language pairs. This means that the results are
not directly comparable across languages. In particular, the sentences tend to have less variety for several
low-resource languages, e.g. "Tom needed water", "Tom needs water", "Tom is getting water", ...

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'validation' 1000
  • Features:
{
    "source_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "source_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    }
}

tatoeba.ara

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/tatoeba.ara')
  • Description:
his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.
For each languages, we have selected 1000 English sentences and their translations, if available. Please check
this paper for a description of the languages, their families and scripts as well as baseline results.
Please note that the English sentences are not identical for all language pairs. This means that the results are
not directly comparable across languages. In particular, the sentences tend to have less variety for several
low-resource languages, e.g. "Tom needed water", "Tom needs water", "Tom is getting water", ...

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'validation' 1000
  • Features:
{
    "source_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "source_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    }
}

tatoeba.ben

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/tatoeba.ben')
  • Description:
his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.
For each languages, we have selected 1000 English sentences and their translations, if available. Please check
this paper for a description of the languages, their families and scripts as well as baseline results.
Please note that the English sentences are not identical for all language pairs. This means that the results are
not directly comparable across languages. In particular, the sentences tend to have less variety for several
low-resource languages, e.g. "Tom needed water", "Tom needs water", "Tom is getting water", ...

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'validation' 1000
  • Features:
{
    "source_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "source_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    }
}

tatoeba.bul

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/tatoeba.bul')
  • Description:
his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.
For each languages, we have selected 1000 English sentences and their translations, if available. Please check
this paper for a description of the languages, their families and scripts as well as baseline results.
Please note that the English sentences are not identical for all language pairs. This means that the results are
not directly comparable across languages. In particular, the sentences tend to have less variety for several
low-resource languages, e.g. "Tom needed water", "Tom needs water", "Tom is getting water", ...

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'validation' 1000
  • Features:
{
    "source_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "source_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    }
}

tatoeba.deu

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/tatoeba.deu')
  • Description:
his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.
For each languages, we have selected 1000 English sentences and their translations, if available. Please check
this paper for a description of the languages, their families and scripts as well as baseline results.
Please note that the English sentences are not identical for all language pairs. This means that the results are
not directly comparable across languages. In particular, the sentences tend to have less variety for several
low-resource languages, e.g. "Tom needed water", "Tom needs water", "Tom is getting water", ...

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'validation' 1000
  • Features:
{
    "source_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "source_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    }
}

tatoeba.cmn

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/tatoeba.cmn')
  • Description:
his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.
For each languages, we have selected 1000 English sentences and their translations, if available. Please check
this paper for a description of the languages, their families and scripts as well as baseline results.
Please note that the English sentences are not identical for all language pairs. This means that the results are
not directly comparable across languages. In particular, the sentences tend to have less variety for several
low-resource languages, e.g. "Tom needed water", "Tom needs water", "Tom is getting water", ...

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'validation' 1000
  • Features:
{
    "source_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "source_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    }
}

tatoeba.ell

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/tatoeba.ell')
  • Description:
his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.
For each languages, we have selected 1000 English sentences and their translations, if available. Please check
this paper for a description of the languages, their families and scripts as well as baseline results.
Please note that the English sentences are not identical for all language pairs. This means that the results are
not directly comparable across languages. In particular, the sentences tend to have less variety for several
low-resource languages, e.g. "Tom needed water", "Tom needs water", "Tom is getting water", ...

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'validation' 1000
  • Features:
{
    "source_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "source_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    }
}

tatoeba.est

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/tatoeba.est')
  • Description:
his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.
For each languages, we have selected 1000 English sentences and their translations, if available. Please check
this paper for a description of the languages, their families and scripts as well as baseline results.
Please note that the English sentences are not identical for all language pairs. This means that the results are
not directly comparable across languages. In particular, the sentences tend to have less variety for several
low-resource languages, e.g. "Tom needed water", "Tom needs water", "Tom is getting water", ...

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'validation' 1000
  • Features:
{
    "source_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "source_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    }
}

tatoeba.eus

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/tatoeba.eus')
  • Description:
his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.
For each languages, we have selected 1000 English sentences and their translations, if available. Please check
this paper for a description of the languages, their families and scripts as well as baseline results.
Please note that the English sentences are not identical for all language pairs. This means that the results are
not directly comparable across languages. In particular, the sentences tend to have less variety for several
low-resource languages, e.g. "Tom needed water", "Tom needs water", "Tom is getting water", ...

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'validation' 1000
  • Features:
{
    "source_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "source_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    }
}

tatoeba.fin

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/tatoeba.fin')
  • Description:
his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.
For each languages, we have selected 1000 English sentences and their translations, if available. Please check
this paper for a description of the languages, their families and scripts as well as baseline results.
Please note that the English sentences are not identical for all language pairs. This means that the results are
not directly comparable across languages. In particular, the sentences tend to have less variety for several
low-resource languages, e.g. "Tom needed water", "Tom needs water", "Tom is getting water", ...

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'validation' 1000
  • Features:
{
    "source_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "source_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    }
}

tatoeba.fra

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/tatoeba.fra')
  • Description:
his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.
For each languages, we have selected 1000 English sentences and their translations, if available. Please check
this paper for a description of the languages, their families and scripts as well as baseline results.
Please note that the English sentences are not identical for all language pairs. This means that the results are
not directly comparable across languages. In particular, the sentences tend to have less variety for several
low-resource languages, e.g. "Tom needed water", "Tom needs water", "Tom is getting water", ...

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'validation' 1000
  • Features:
{
    "source_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "source_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    }
}

tatoeba.heb

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/tatoeba.heb')
  • Description:
his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.
For each languages, we have selected 1000 English sentences and their translations, if available. Please check
this paper for a description of the languages, their families and scripts as well as baseline results.
Please note that the English sentences are not identical for all language pairs. This means that the results are
not directly comparable across languages. In particular, the sentences tend to have less variety for several
low-resource languages, e.g. "Tom needed water", "Tom needs water", "Tom is getting water", ...

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'validation' 1000
  • Features:
{
    "source_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "source_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    }
}

tatoeba.hin

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/tatoeba.hin')
  • Description:
his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.
For each languages, we have selected 1000 English sentences and their translations, if available. Please check
this paper for a description of the languages, their families and scripts as well as baseline results.
Please note that the English sentences are not identical for all language pairs. This means that the results are
not directly comparable across languages. In particular, the sentences tend to have less variety for several
low-resource languages, e.g. "Tom needed water", "Tom needs water", "Tom is getting water", ...

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'validation' 1000
  • Features:
{
    "source_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "source_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    }
}

tatoeba.hun

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/tatoeba.hun')
  • Description:
his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.
For each languages, we have selected 1000 English sentences and their translations, if available. Please check
this paper for a description of the languages, their families and scripts as well as baseline results.
Please note that the English sentences are not identical for all language pairs. This means that the results are
not directly comparable across languages. In particular, the sentences tend to have less variety for several
low-resource languages, e.g. "Tom needed water", "Tom needs water", "Tom is getting water", ...

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'validation' 1000
  • Features:
{
    "source_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "source_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    }
}

tatoeba.ind

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/tatoeba.ind')
  • Description:
his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.
For each languages, we have selected 1000 English sentences and their translations, if available. Please check
this paper for a description of the languages, their families and scripts as well as baseline results.
Please note that the English sentences are not identical for all language pairs. This means that the results are
not directly comparable across languages. In particular, the sentences tend to have less variety for several
low-resource languages, e.g. "Tom needed water", "Tom needs water", "Tom is getting water", ...

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'validation' 1000
  • Features:
{
    "source_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "source_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    }
}

tatoeba.ita

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/tatoeba.ita')
  • Description:
his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.
For each languages, we have selected 1000 English sentences and their translations, if available. Please check
this paper for a description of the languages, their families and scripts as well as baseline results.
Please note that the English sentences are not identical for all language pairs. This means that the results are
not directly comparable across languages. In particular, the sentences tend to have less variety for several
low-resource languages, e.g. "Tom needed water", "Tom needs water", "Tom is getting water", ...

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'validation' 1000
  • Features:
{
    "source_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "source_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    }
}

tatoeba.jav

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/tatoeba.jav')
  • Description:
his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.
For each languages, we have selected 1000 English sentences and their translations, if available. Please check
this paper for a description of the languages, their families and scripts as well as baseline results.
Please note that the English sentences are not identical for all language pairs. This means that the results are
not directly comparable across languages. In particular, the sentences tend to have less variety for several
low-resource languages, e.g. "Tom needed water", "Tom needs water", "Tom is getting water", ...

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'validation' 205
  • Features:
{
    "source_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "source_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    }
}

tatoeba.jpn

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/tatoeba.jpn')
  • Description:
his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.
For each languages, we have selected 1000 English sentences and their translations, if available. Please check
this paper for a description of the languages, their families and scripts as well as baseline results.
Please note that the English sentences are not identical for all language pairs. This means that the results are
not directly comparable across languages. In particular, the sentences tend to have less variety for several
low-resource languages, e.g. "Tom needed water", "Tom needs water", "Tom is getting water", ...

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'validation' 1000
  • Features:
{
    "source_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "source_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    }
}

tatoeba.kat

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/tatoeba.kat')
  • Description:
his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.
For each languages, we have selected 1000 English sentences and their translations, if available. Please check
this paper for a description of the languages, their families and scripts as well as baseline results.
Please note that the English sentences are not identical for all language pairs. This means that the results are
not directly comparable across languages. In particular, the sentences tend to have less variety for several
low-resource languages, e.g. "Tom needed water", "Tom needs water", "Tom is getting water", ...

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'validation' 746
  • Features:
{
    "source_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "source_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    }
}

tatoeba.kaz

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/tatoeba.kaz')
  • Description:
his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.
For each languages, we have selected 1000 English sentences and their translations, if available. Please check
this paper for a description of the languages, their families and scripts as well as baseline results.
Please note that the English sentences are not identical for all language pairs. This means that the results are
not directly comparable across languages. In particular, the sentences tend to have less variety for several
low-resource languages, e.g. "Tom needed water", "Tom needs water", "Tom is getting water", ...

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'validation' 575
  • Features:
{
    "source_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "source_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    }
}

tatoeba.kor

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/tatoeba.kor')
  • Description:
his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.
For each languages, we have selected 1000 English sentences and their translations, if available. Please check
this paper for a description of the languages, their families and scripts as well as baseline results.
Please note that the English sentences are not identical for all language pairs. This means that the results are
not directly comparable across languages. In particular, the sentences tend to have less variety for several
low-resource languages, e.g. "Tom needed water", "Tom needs water", "Tom is getting water", ...

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'validation' 1000
  • Features:
{
    "source_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "source_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    }
}

tatoeba.mal

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/tatoeba.mal')
  • Description:
his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.
For each languages, we have selected 1000 English sentences and their translations, if available. Please check
this paper for a description of the languages, their families and scripts as well as baseline results.
Please note that the English sentences are not identical for all language pairs. This means that the results are
not directly comparable across languages. In particular, the sentences tend to have less variety for several
low-resource languages, e.g. "Tom needed water", "Tom needs water", "Tom is getting water", ...

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'validation' 687
  • Features:
{
    "source_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "source_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    }
}

tatoeba.mar

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/tatoeba.mar')
  • Description:
his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.
For each languages, we have selected 1000 English sentences and their translations, if available. Please check
this paper for a description of the languages, their families and scripts as well as baseline results.
Please note that the English sentences are not identical for all language pairs. This means that the results are
not directly comparable across languages. In particular, the sentences tend to have less variety for several
low-resource languages, e.g. "Tom needed water", "Tom needs water", "Tom is getting water", ...

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'validation' 1000
  • Features:
{
    "source_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "source_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    }
}

tatoeba.nld

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/tatoeba.nld')
  • Description:
his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.
For each languages, we have selected 1000 English sentences and their translations, if available. Please check
this paper for a description of the languages, their families and scripts as well as baseline results.
Please note that the English sentences are not identical for all language pairs. This means that the results are
not directly comparable across languages. In particular, the sentences tend to have less variety for several
low-resource languages, e.g. "Tom needed water", "Tom needs water", "Tom is getting water", ...

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'validation' 1000
  • Features:
{
    "source_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "source_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    }
}

tatoeba.pes

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/tatoeba.pes')
  • Description:
his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.
For each languages, we have selected 1000 English sentences and their translations, if available. Please check
this paper for a description of the languages, their families and scripts as well as baseline results.
Please note that the English sentences are not identical for all language pairs. This means that the results are
not directly comparable across languages. In particular, the sentences tend to have less variety for several
low-resource languages, e.g. "Tom needed water", "Tom needs water", "Tom is getting water", ...

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'validation' 1000
  • Features:
{
    "source_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "source_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    }
}

tatoeba.por

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/tatoeba.por')
  • Description:
his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.
For each languages, we have selected 1000 English sentences and their translations, if available. Please check
this paper for a description of the languages, their families and scripts as well as baseline results.
Please note that the English sentences are not identical for all language pairs. This means that the results are
not directly comparable across languages. In particular, the sentences tend to have less variety for several
low-resource languages, e.g. "Tom needed water", "Tom needs water", "Tom is getting water", ...

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'validation' 1000
  • Features:
{
    "source_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "source_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    }
}

tatoeba.rus

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/tatoeba.rus')
  • Description:
his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.
For each languages, we have selected 1000 English sentences and their translations, if available. Please check
this paper for a description of the languages, their families and scripts as well as baseline results.
Please note that the English sentences are not identical for all language pairs. This means that the results are
not directly comparable across languages. In particular, the sentences tend to have less variety for several
low-resource languages, e.g. "Tom needed water", "Tom needs water", "Tom is getting water", ...

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'validation' 1000
  • Features:
{
    "source_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "source_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    }
}

tatoeba.spa

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/tatoeba.spa')
  • Description:
his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.
For each languages, we have selected 1000 English sentences and their translations, if available. Please check
this paper for a description of the languages, their families and scripts as well as baseline results.
Please note that the English sentences are not identical for all language pairs. This means that the results are
not directly comparable across languages. In particular, the sentences tend to have less variety for several
low-resource languages, e.g. "Tom needed water", "Tom needs water", "Tom is getting water", ...

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'validation' 1000
  • Features:
{
    "source_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "source_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    }
}

tatoeba.swh

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/tatoeba.swh')
  • Description:
his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.
For each languages, we have selected 1000 English sentences and their translations, if available. Please check
this paper for a description of the languages, their families and scripts as well as baseline results.
Please note that the English sentences are not identical for all language pairs. This means that the results are
not directly comparable across languages. In particular, the sentences tend to have less variety for several
low-resource languages, e.g. "Tom needed water", "Tom needs water", "Tom is getting water", ...

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'validation' 390
  • Features:
{
    "source_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "source_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    }
}

tatoeba.tam

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/tatoeba.tam')
  • Description:
his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.
For each languages, we have selected 1000 English sentences and their translations, if available. Please check
this paper for a description of the languages, their families and scripts as well as baseline results.
Please note that the English sentences are not identical for all language pairs. This means that the results are
not directly comparable across languages. In particular, the sentences tend to have less variety for several
low-resource languages, e.g. "Tom needed water", "Tom needs water", "Tom is getting water", ...

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'validation' 307
  • Features:
{
    "source_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "source_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    }
}

tatoeba.tel

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/tatoeba.tel')
  • Description:
his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.
For each languages, we have selected 1000 English sentences and their translations, if available. Please check
this paper for a description of the languages, their families and scripts as well as baseline results.
Please note that the English sentences are not identical for all language pairs. This means that the results are
not directly comparable across languages. In particular, the sentences tend to have less variety for several
low-resource languages, e.g. "Tom needed water", "Tom needs water", "Tom is getting water", ...

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'validation' 234
  • Features:
{
    "source_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "source_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    }
}

tatoeba.tgl

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/tatoeba.tgl')
  • Description:
his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.
For each languages, we have selected 1000 English sentences and their translations, if available. Please check
this paper for a description of the languages, their families and scripts as well as baseline results.
Please note that the English sentences are not identical for all language pairs. This means that the results are
not directly comparable across languages. In particular, the sentences tend to have less variety for several
low-resource languages, e.g. "Tom needed water", "Tom needs water", "Tom is getting water", ...

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'validation' 1000
  • Features:
{
    "source_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "source_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    }
}

tatoeba.tha

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/tatoeba.tha')
  • Description:
his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.
For each languages, we have selected 1000 English sentences and their translations, if available. Please check
this paper for a description of the languages, their families and scripts as well as baseline results.
Please note that the English sentences are not identical for all language pairs. This means that the results are
not directly comparable across languages. In particular, the sentences tend to have less variety for several
low-resource languages, e.g. "Tom needed water", "Tom needs water", "Tom is getting water", ...

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'validation' 548
  • Features:
{
    "source_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "source_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    }
}

tatoeba.tur

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/tatoeba.tur')
  • Description:
his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.
For each languages, we have selected 1000 English sentences and their translations, if available. Please check
this paper for a description of the languages, their families and scripts as well as baseline results.
Please note that the English sentences are not identical for all language pairs. This means that the results are
not directly comparable across languages. In particular, the sentences tend to have less variety for several
low-resource languages, e.g. "Tom needed water", "Tom needs water", "Tom is getting water", ...

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'validation' 1000
  • Features:
{
    "source_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "source_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    }
}

tatoeba.urd

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/tatoeba.urd')
  • Description:
his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.
For each languages, we have selected 1000 English sentences and their translations, if available. Please check
this paper for a description of the languages, their families and scripts as well as baseline results.
Please note that the English sentences are not identical for all language pairs. This means that the results are
not directly comparable across languages. In particular, the sentences tend to have less variety for several
low-resource languages, e.g. "Tom needed water", "Tom needs water", "Tom is getting water", ...

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'validation' 1000
  • Features:
{
    "source_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "source_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    }
}

tatoeba.vie

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/tatoeba.vie')
  • Description:
his data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.
For each languages, we have selected 1000 English sentences and their translations, if available. Please check
this paper for a description of the languages, their families and scripts as well as baseline results.
Please note that the English sentences are not identical for all language pairs. This means that the results are
not directly comparable across languages. In particular, the sentences tend to have less variety for several
low-resource languages, e.g. "Tom needed water", "Tom needs water", "Tom is getting water", ...

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'validation' 1000
  • Features:
{
    "source_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_sentence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "source_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "target_lang": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    }
}

udpos.Afrikaans

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/udpos.Afrikaans')
  • Description:
Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological
features, and syntactic dependencies) across different human languages. UD is an open community effort with over 200
contributors producing more than 100 treebanks in over 70 languages. If you’re new to UD, you should start by reading
the first part of the Short Introduction and then browsing the annotation guidelines.

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 425
'train' 1315
'validation' 194
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "pos_tags": {
        "feature": {
            "num_classes": 17,
            "names": [
                "ADJ",
                "ADP",
                "ADV",
                "AUX",
                "CCONJ",
                "DET",
                "INTJ",
                "NOUN",
                "NUM",
                "PART",
                "PRON",
                "PROPN",
                "PUNCT",
                "SCONJ",
                "SYM",
                "VERB",
                "X"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

udpos.Arabic

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/udpos.Arabic')
  • Description:
Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological
features, and syntactic dependencies) across different human languages. UD is an open community effort with over 200
contributors producing more than 100 treebanks in over 70 languages. If you’re new to UD, you should start by reading
the first part of the Short Introduction and then browsing the annotation guidelines.

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 1680
'train' 6075
'validation' 909
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "pos_tags": {
        "feature": {
            "num_classes": 17,
            "names": [
                "ADJ",
                "ADP",
                "ADV",
                "AUX",
                "CCONJ",
                "DET",
                "INTJ",
                "NOUN",
                "NUM",
                "PART",
                "PRON",
                "PROPN",
                "PUNCT",
                "SCONJ",
                "SYM",
                "VERB",
                "X"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

udpos.Basque

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/udpos.Basque')
  • Description:
Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological
features, and syntactic dependencies) across different human languages. UD is an open community effort with over 200
contributors producing more than 100 treebanks in over 70 languages. If you’re new to UD, you should start by reading
the first part of the Short Introduction and then browsing the annotation guidelines.

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 1799
'train' 5396
'validation' 1798
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "pos_tags": {
        "feature": {
            "num_classes": 17,
            "names": [
                "ADJ",
                "ADP",
                "ADV",
                "AUX",
                "CCONJ",
                "DET",
                "INTJ",
                "NOUN",
                "NUM",
                "PART",
                "PRON",
                "PROPN",
                "PUNCT",
                "SCONJ",
                "SYM",
                "VERB",
                "X"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

udpos.Bulgarian

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/udpos.Bulgarian')
  • Description:
Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological
features, and syntactic dependencies) across different human languages. UD is an open community effort with over 200
contributors producing more than 100 treebanks in over 70 languages. If you’re new to UD, you should start by reading
the first part of the Short Introduction and then browsing the annotation guidelines.

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 1116
'train' 8907
'validation' 1115
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "pos_tags": {
        "feature": {
            "num_classes": 17,
            "names": [
                "ADJ",
                "ADP",
                "ADV",
                "AUX",
                "CCONJ",
                "DET",
                "INTJ",
                "NOUN",
                "NUM",
                "PART",
                "PRON",
                "PROPN",
                "PUNCT",
                "SCONJ",
                "SYM",
                "VERB",
                "X"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

udpos.Dutch

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/udpos.Dutch')
  • Description:
Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological
features, and syntactic dependencies) across different human languages. UD is an open community effort with over 200
contributors producing more than 100 treebanks in over 70 languages. If you’re new to UD, you should start by reading
the first part of the Short Introduction and then browsing the annotation guidelines.

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 1471
'train' 18051
'validation' 1394
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "pos_tags": {
        "feature": {
            "num_classes": 17,
            "names": [
                "ADJ",
                "ADP",
                "ADV",
                "AUX",
                "CCONJ",
                "DET",
                "INTJ",
                "NOUN",
                "NUM",
                "PART",
                "PRON",
                "PROPN",
                "PUNCT",
                "SCONJ",
                "SYM",
                "VERB",
                "X"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

udpos.English

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/udpos.English')
  • Description:
Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological
features, and syntactic dependencies) across different human languages. UD is an open community effort with over 200
contributors producing more than 100 treebanks in over 70 languages. If you’re new to UD, you should start by reading
the first part of the Short Introduction and then browsing the annotation guidelines.

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 5440
'train' 21253
'validation' 3974
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "pos_tags": {
        "feature": {
            "num_classes": 17,
            "names": [
                "ADJ",
                "ADP",
                "ADV",
                "AUX",
                "CCONJ",
                "DET",
                "INTJ",
                "NOUN",
                "NUM",
                "PART",
                "PRON",
                "PROPN",
                "PUNCT",
                "SCONJ",
                "SYM",
                "VERB",
                "X"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

udpos.Estonian

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/udpos.Estonian')
  • Description:
Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological
features, and syntactic dependencies) across different human languages. UD is an open community effort with over 200
contributors producing more than 100 treebanks in over 70 languages. If you’re new to UD, you should start by reading
the first part of the Short Introduction and then browsing the annotation guidelines.

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 3760
'train' 25749
'validation' 3125
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "pos_tags": {
        "feature": {
            "num_classes": 17,
            "names": [
                "ADJ",
                "ADP",
                "ADV",
                "AUX",
                "CCONJ",
                "DET",
                "INTJ",
                "NOUN",
                "NUM",
                "PART",
                "PRON",
                "PROPN",
                "PUNCT",
                "SCONJ",
                "SYM",
                "VERB",
                "X"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

udpos.Finnish

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/udpos.Finnish')
  • Description:
Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological
features, and syntactic dependencies) across different human languages. UD is an open community effort with over 200
contributors producing more than 100 treebanks in over 70 languages. If you’re new to UD, you should start by reading
the first part of the Short Introduction and then browsing the annotation guidelines.

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 4422
'train' 27198
'validation' 3239
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "pos_tags": {
        "feature": {
            "num_classes": 17,
            "names": [
                "ADJ",
                "ADP",
                "ADV",
                "AUX",
                "CCONJ",
                "DET",
                "INTJ",
                "NOUN",
                "NUM",
                "PART",
                "PRON",
                "PROPN",
                "PUNCT",
                "SCONJ",
                "SYM",
                "VERB",
                "X"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

udpos.French

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/udpos.French')
  • Description:
Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological
features, and syntactic dependencies) across different human languages. UD is an open community effort with over 200
contributors producing more than 100 treebanks in over 70 languages. If you’re new to UD, you should start by reading
the first part of the Short Introduction and then browsing the annotation guidelines.

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 9465
'train' 47308
'validation' 5979
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "pos_tags": {
        "feature": {
            "num_classes": 17,
            "names": [
                "ADJ",
                "ADP",
                "ADV",
                "AUX",
                "CCONJ",
                "DET",
                "INTJ",
                "NOUN",
                "NUM",
                "PART",
                "PRON",
                "PROPN",
                "PUNCT",
                "SCONJ",
                "SYM",
                "VERB",
                "X"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

udpos.German

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/udpos.German')
  • Description:
Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological
features, and syntactic dependencies) across different human languages. UD is an open community effort with over 200
contributors producing more than 100 treebanks in over 70 languages. If you’re new to UD, you should start by reading
the first part of the Short Introduction and then browsing the annotation guidelines.

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 22458
'train' 166849
'validation' 19233
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "pos_tags": {
        "feature": {
            "num_classes": 17,
            "names": [
                "ADJ",
                "ADP",
                "ADV",
                "AUX",
                "CCONJ",
                "DET",
                "INTJ",
                "NOUN",
                "NUM",
                "PART",
                "PRON",
                "PROPN",
                "PUNCT",
                "SCONJ",
                "SYM",
                "VERB",
                "X"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

udpos.Greek

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/udpos.Greek')
  • Description:
Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological
features, and syntactic dependencies) across different human languages. UD is an open community effort with over 200
contributors producing more than 100 treebanks in over 70 languages. If you’re new to UD, you should start by reading
the first part of the Short Introduction and then browsing the annotation guidelines.

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 2809
'train' 28152
'validation' 2559
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "pos_tags": {
        "feature": {
            "num_classes": 17,
            "names": [
                "ADJ",
                "ADP",
                "ADV",
                "AUX",
                "CCONJ",
                "DET",
                "INTJ",
                "NOUN",
                "NUM",
                "PART",
                "PRON",
                "PROPN",
                "PUNCT",
                "SCONJ",
                "SYM",
                "VERB",
                "X"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

udpos.Hebrew

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/udpos.Hebrew')
  • Description:
Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological
features, and syntactic dependencies) across different human languages. UD is an open community effort with over 200
contributors producing more than 100 treebanks in over 70 languages. If you’re new to UD, you should start by reading
the first part of the Short Introduction and then browsing the annotation guidelines.

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 491
'train' 5241
'validation' 484
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "pos_tags": {
        "feature": {
            "num_classes": 17,
            "names": [
                "ADJ",
                "ADP",
                "ADV",
                "AUX",
                "CCONJ",
                "DET",
                "INTJ",
                "NOUN",
                "NUM",
                "PART",
                "PRON",
                "PROPN",
                "PUNCT",
                "SCONJ",
                "SYM",
                "VERB",
                "X"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

udpos.Hindi

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/udpos.Hindi')
  • Description:
Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological
features, and syntactic dependencies) across different human languages. UD is an open community effort with over 200
contributors producing more than 100 treebanks in over 70 languages. If you’re new to UD, you should start by reading
the first part of the Short Introduction and then browsing the annotation guidelines.

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 2684
'train' 13304
'validation' 1659
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "pos_tags": {
        "feature": {
            "num_classes": 17,
            "names": [
                "ADJ",
                "ADP",
                "ADV",
                "AUX",
                "CCONJ",
                "DET",
                "INTJ",
                "NOUN",
                "NUM",
                "PART",
                "PRON",
                "PROPN",
                "PUNCT",
                "SCONJ",
                "SYM",
                "VERB",
                "X"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

udpos.Hungarian

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/udpos.Hungarian')
  • Description:
Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological
features, and syntactic dependencies) across different human languages. UD is an open community effort with over 200
contributors producing more than 100 treebanks in over 70 languages. If you’re new to UD, you should start by reading
the first part of the Short Introduction and then browsing the annotation guidelines.

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 449
'train' 910
'validation' 441
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "pos_tags": {
        "feature": {
            "num_classes": 17,
            "names": [
                "ADJ",
                "ADP",
                "ADV",
                "AUX",
                "CCONJ",
                "DET",
                "INTJ",
                "NOUN",
                "NUM",
                "PART",
                "PRON",
                "PROPN",
                "PUNCT",
                "SCONJ",
                "SYM",
                "VERB",
                "X"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

udpos.Indonesian

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/udpos.Indonesian')
  • Description:
Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological
features, and syntactic dependencies) across different human languages. UD is an open community effort with over 200
contributors producing more than 100 treebanks in over 70 languages. If you’re new to UD, you should start by reading
the first part of the Short Introduction and then browsing the annotation guidelines.

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 1557
'train' 4477
'validation' 559
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "pos_tags": {
        "feature": {
            "num_classes": 17,
            "names": [
                "ADJ",
                "ADP",
                "ADV",
                "AUX",
                "CCONJ",
                "DET",
                "INTJ",
                "NOUN",
                "NUM",
                "PART",
                "PRON",
                "PROPN",
                "PUNCT",
                "SCONJ",
                "SYM",
                "VERB",
                "X"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

udpos.Italian

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/udpos.Italian')
  • Description:
Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological
features, and syntactic dependencies) across different human languages. UD is an open community effort with over 200
contributors producing more than 100 treebanks in over 70 languages. If you’re new to UD, you should start by reading
the first part of the Short Introduction and then browsing the annotation guidelines.

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 3518
'train' 29685
'validation' 2278
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "pos_tags": {
        "feature": {
            "num_classes": 17,
            "names": [
                "ADJ",
                "ADP",
                "ADV",
                "AUX",
                "CCONJ",
                "DET",
                "INTJ",
                "NOUN",
                "NUM",
                "PART",
                "PRON",
                "PROPN",
                "PUNCT",
                "SCONJ",
                "SYM",
                "VERB",
                "X"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

udpos.Japanese

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/udpos.Japanese')
  • Description:
Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological
features, and syntactic dependencies) across different human languages. UD is an open community effort with over 200
contributors producing more than 100 treebanks in over 70 languages. If you’re new to UD, you should start by reading
the first part of the Short Introduction and then browsing the annotation guidelines.

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 2372
'train' 7125
'validation' 511
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "pos_tags": {
        "feature": {
            "num_classes": 17,
            "names": [
                "ADJ",
                "ADP",
                "ADV",
                "AUX",
                "CCONJ",
                "DET",
                "INTJ",
                "NOUN",
                "NUM",
                "PART",
                "PRON",
                "PROPN",
                "PUNCT",
                "SCONJ",
                "SYM",
                "VERB",
                "X"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

udpos.Kazakh

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/udpos.Kazakh')
  • Description:
Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological
features, and syntactic dependencies) across different human languages. UD is an open community effort with over 200
contributors producing more than 100 treebanks in over 70 languages. If you’re new to UD, you should start by reading
the first part of the Short Introduction and then browsing the annotation guidelines.

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 1047
'train' 31
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "pos_tags": {
        "feature": {
            "num_classes": 17,
            "names": [
                "ADJ",
                "ADP",
                "ADV",
                "AUX",
                "CCONJ",
                "DET",
                "INTJ",
                "NOUN",
                "NUM",
                "PART",
                "PRON",
                "PROPN",
                "PUNCT",
                "SCONJ",
                "SYM",
                "VERB",
                "X"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

udpos.Korean

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/udpos.Korean')
  • Description:
Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological
features, and syntactic dependencies) across different human languages. UD is an open community effort with over 200
contributors producing more than 100 treebanks in over 70 languages. If you’re new to UD, you should start by reading
the first part of the Short Introduction and then browsing the annotation guidelines.

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 4276
'train' 27410
'validation' 3016
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "pos_tags": {
        "feature": {
            "num_classes": 17,
            "names": [
                "ADJ",
                "ADP",
                "ADV",
                "AUX",
                "CCONJ",
                "DET",
                "INTJ",
                "NOUN",
                "NUM",
                "PART",
                "PRON",
                "PROPN",
                "PUNCT",
                "SCONJ",
                "SYM",
                "VERB",
                "X"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

udpos.Chinese

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/udpos.Chinese')
  • Description:
Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological
features, and syntactic dependencies) across different human languages. UD is an open community effort with over 200
contributors producing more than 100 treebanks in over 70 languages. If you’re new to UD, you should start by reading
the first part of the Short Introduction and then browsing the annotation guidelines.

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 5528
'train' 18998
'validation' 3038
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "pos_tags": {
        "feature": {
            "num_classes": 17,
            "names": [
                "ADJ",
                "ADP",
                "ADV",
                "AUX",
                "CCONJ",
                "DET",
                "INTJ",
                "NOUN",
                "NUM",
                "PART",
                "PRON",
                "PROPN",
                "PUNCT",
                "SCONJ",
                "SYM",
                "VERB",
                "X"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

udpos.Marathi

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/udpos.Marathi')
  • Description:
Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological
features, and syntactic dependencies) across different human languages. UD is an open community effort with over 200
contributors producing more than 100 treebanks in over 70 languages. If you’re new to UD, you should start by reading
the first part of the Short Introduction and then browsing the annotation guidelines.

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 47
'train' 373
'validation' 46
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "pos_tags": {
        "feature": {
            "num_classes": 17,
            "names": [
                "ADJ",
                "ADP",
                "ADV",
                "AUX",
                "CCONJ",
                "DET",
                "INTJ",
                "NOUN",
                "NUM",
                "PART",
                "PRON",
                "PROPN",
                "PUNCT",
                "SCONJ",
                "SYM",
                "VERB",
                "X"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

udpos.Persian

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/udpos.Persian')
  • Description:
Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological
features, and syntactic dependencies) across different human languages. UD is an open community effort with over 200
contributors producing more than 100 treebanks in over 70 languages. If you’re new to UD, you should start by reading
the first part of the Short Introduction and then browsing the annotation guidelines.

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 600
'train' 4798
'validation' 599
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "pos_tags": {
        "feature": {
            "num_classes": 17,
            "names": [
                "ADJ",
                "ADP",
                "ADV",
                "AUX",
                "CCONJ",
                "DET",
                "INTJ",
                "NOUN",
                "NUM",
                "PART",
                "PRON",
                "PROPN",
                "PUNCT",
                "SCONJ",
                "SYM",
                "VERB",
                "X"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

udpos.Portuguese

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/udpos.Portuguese')
  • Description:
Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological
features, and syntactic dependencies) across different human languages. UD is an open community effort with over 200
contributors producing more than 100 treebanks in over 70 languages. If you’re new to UD, you should start by reading
the first part of the Short Introduction and then browsing the annotation guidelines.

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 2681
'train' 17992
'validation' 1770
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "pos_tags": {
        "feature": {
            "num_classes": 17,
            "names": [
                "ADJ",
                "ADP",
                "ADV",
                "AUX",
                "CCONJ",
                "DET",
                "INTJ",
                "NOUN",
                "NUM",
                "PART",
                "PRON",
                "PROPN",
                "PUNCT",
                "SCONJ",
                "SYM",
                "VERB",
                "X"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

udpos.Russian

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/udpos.Russian')
  • Description:
Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological
features, and syntactic dependencies) across different human languages. UD is an open community effort with over 200
contributors producing more than 100 treebanks in over 70 languages. If you’re new to UD, you should start by reading
the first part of the Short Introduction and then browsing the annotation guidelines.

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 11336
'train' 67435
'validation' 9960
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "pos_tags": {
        "feature": {
            "num_classes": 17,
            "names": [
                "ADJ",
                "ADP",
                "ADV",
                "AUX",
                "CCONJ",
                "DET",
                "INTJ",
                "NOUN",
                "NUM",
                "PART",
                "PRON",
                "PROPN",
                "PUNCT",
                "SCONJ",
                "SYM",
                "VERB",
                "X"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

udpos.Spanish

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/udpos.Spanish')
  • Description:
Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological
features, and syntactic dependencies) across different human languages. UD is an open community effort with over 200
contributors producing more than 100 treebanks in over 70 languages. If you’re new to UD, you should start by reading
the first part of the Short Introduction and then browsing the annotation guidelines.

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 3147
'train' 28492
'validation' 3054
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "pos_tags": {
        "feature": {
            "num_classes": 17,
            "names": [
                "ADJ",
                "ADP",
                "ADV",
                "AUX",
                "CCONJ",
                "DET",
                "INTJ",
                "NOUN",
                "NUM",
                "PART",
                "PRON",
                "PROPN",
                "PUNCT",
                "SCONJ",
                "SYM",
                "VERB",
                "X"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

udpos.Tagalog

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/udpos.Tagalog')
  • Description:
Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological
features, and syntactic dependencies) across different human languages. UD is an open community effort with over 200
contributors producing more than 100 treebanks in over 70 languages. If you’re new to UD, you should start by reading
the first part of the Short Introduction and then browsing the annotation guidelines.

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 55
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "pos_tags": {
        "feature": {
            "num_classes": 17,
            "names": [
                "ADJ",
                "ADP",
                "ADV",
                "AUX",
                "CCONJ",
                "DET",
                "INTJ",
                "NOUN",
                "NUM",
                "PART",
                "PRON",
                "PROPN",
                "PUNCT",
                "SCONJ",
                "SYM",
                "VERB",
                "X"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

udpos.Tamil

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/udpos.Tamil')
  • Description:
Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological
features, and syntactic dependencies) across different human languages. UD is an open community effort with over 200
contributors producing more than 100 treebanks in over 70 languages. If you’re new to UD, you should start by reading
the first part of the Short Introduction and then browsing the annotation guidelines.

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 120
'train' 400
'validation' 80
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "pos_tags": {
        "feature": {
            "num_classes": 17,
            "names": [
                "ADJ",
                "ADP",
                "ADV",
                "AUX",
                "CCONJ",
                "DET",
                "INTJ",
                "NOUN",
                "NUM",
                "PART",
                "PRON",
                "PROPN",
                "PUNCT",
                "SCONJ",
                "SYM",
                "VERB",
                "X"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

udpos.Telugu

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/udpos.Telugu')
  • Description:
Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological
features, and syntactic dependencies) across different human languages. UD is an open community effort with over 200
contributors producing more than 100 treebanks in over 70 languages. If you’re new to UD, you should start by reading
the first part of the Short Introduction and then browsing the annotation guidelines.

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 146
'train' 1051
'validation' 131
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "pos_tags": {
        "feature": {
            "num_classes": 17,
            "names": [
                "ADJ",
                "ADP",
                "ADV",
                "AUX",
                "CCONJ",
                "DET",
                "INTJ",
                "NOUN",
                "NUM",
                "PART",
                "PRON",
                "PROPN",
                "PUNCT",
                "SCONJ",
                "SYM",
                "VERB",
                "X"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

udpos.Thai

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/udpos.Thai')
  • Description:
Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological
features, and syntactic dependencies) across different human languages. UD is an open community effort with over 200
contributors producing more than 100 treebanks in over 70 languages. If you’re new to UD, you should start by reading
the first part of the Short Introduction and then browsing the annotation guidelines.

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 1000
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "pos_tags": {
        "feature": {
            "num_classes": 17,
            "names": [
                "ADJ",
                "ADP",
                "ADV",
                "AUX",
                "CCONJ",
                "DET",
                "INTJ",
                "NOUN",
                "NUM",
                "PART",
                "PRON",
                "PROPN",
                "PUNCT",
                "SCONJ",
                "SYM",
                "VERB",
                "X"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

udpos.Turkish

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/udpos.Turkish')
  • Description:
Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological
features, and syntactic dependencies) across different human languages. UD is an open community effort with over 200
contributors producing more than 100 treebanks in over 70 languages. If you’re new to UD, you should start by reading
the first part of the Short Introduction and then browsing the annotation guidelines.

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 4785
'train' 3664
'validation' 988
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "pos_tags": {
        "feature": {
            "num_classes": 17,
            "names": [
                "ADJ",
                "ADP",
                "ADV",
                "AUX",
                "CCONJ",
                "DET",
                "INTJ",
                "NOUN",
                "NUM",
                "PART",
                "PRON",
                "PROPN",
                "PUNCT",
                "SCONJ",
                "SYM",
                "VERB",
                "X"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

udpos.Urdu

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/udpos.Urdu')
  • Description:
Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological
features, and syntactic dependencies) across different human languages. UD is an open community effort with over 200
contributors producing more than 100 treebanks in over 70 languages. If you’re new to UD, you should start by reading
the first part of the Short Introduction and then browsing the annotation guidelines.

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 535
'train' 4043
'validation' 552
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "pos_tags": {
        "feature": {
            "num_classes": 17,
            "names": [
                "ADJ",
                "ADP",
                "ADV",
                "AUX",
                "CCONJ",
                "DET",
                "INTJ",
                "NOUN",
                "NUM",
                "PART",
                "PRON",
                "PROPN",
                "PUNCT",
                "SCONJ",
                "SYM",
                "VERB",
                "X"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

udpos.Vietnamese

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/udpos.Vietnamese')
  • Description:
Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological
features, and syntactic dependencies) across different human languages. UD is an open community effort with over 200
contributors producing more than 100 treebanks in over 70 languages. If you’re new to UD, you should start by reading
the first part of the Short Introduction and then browsing the annotation guidelines.

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 800
'train' 1400
'validation' 800
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "pos_tags": {
        "feature": {
            "num_classes": 17,
            "names": [
                "ADJ",
                "ADP",
                "ADV",
                "AUX",
                "CCONJ",
                "DET",
                "INTJ",
                "NOUN",
                "NUM",
                "PART",
                "PRON",
                "PROPN",
                "PUNCT",
                "SCONJ",
                "SYM",
                "VERB",
                "X"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

udpos.Yoruba

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:xtreme/udpos.Yoruba')
  • Description:
Universal Dependencies (UD) is a framework for consistent annotation of grammar (parts of speech, morphological
features, and syntactic dependencies) across different human languages. UD is an open community effort with over 200
contributors producing more than 100 treebanks in over 70 languages. If you’re new to UD, you should start by reading
the first part of the Short Introduction and then browsing the annotation guidelines.

The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is a benchmark for the evaluation of
the cross-lingual generalization ability of pre-trained multilingual models. It covers 40 typologically diverse languages
(spanning 12 language families) and includes nine tasks that collectively require reasoning about different levels of
syntax and semantics. The languages in XTREME are selected to maximize language diversity, coverage in existing tasks,
and availability of training data. Among these are many under-studied languages, such as the Dravidian languages Tamil
(spoken in southern India, Sri Lanka, and Singapore), Telugu and Malayalam (spoken mainly in southern India), and the
Niger-Congo languages Swahili and Yoruba, spoken in Africa.
  • License: No known license
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 100
  • Features:
{
    "tokens": {
        "feature": {
            "dtype": "string",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "pos_tags": {
        "feature": {
            "num_classes": 17,
            "names": [
                "ADJ",
                "ADP",
                "ADV",
                "AUX",
                "CCONJ",
                "DET",
                "INTJ",
                "NOUN",
                "NUM",
                "PART",
                "PRON",
                "PROPN",
                "PUNCT",
                "SCONJ",
                "SYM",
                "VERB",
                "X"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}