natural_instructions

• Description:

Description

A compilation of 1600+ tasks phrased as natural instructions. The original task collection can be found at: https://github.com/allenai/natural-instructions No preprocessing or changes were made to this original version.

Note that users of this task collection should consult the underlying licenses of the contained datasets, and cite them accordingly.

• Homepage: https://github.com/allenai/natural-instructions

• Source code: tfds.datasets.natural_instructions.Builder

• Versions:

  • 1.0.0: Initial release.
  • 1.0.1 (default): Added task name field, and fixed ID used for shuffling to use stable IDs.

• Download size: 3.08 GiB

• Dataset size: 4.73 GiB

• Auto-cached (documentation): No

• Splits:

Split	Examples
'train'	5,040,134

• Feature structure:

FeaturesDict({
    'definition': Text(shape=(), dtype=string),
    'id': Text(shape=(), dtype=string),
    'input': Text(shape=(), dtype=string),
    'output': Text(shape=(), dtype=string),
    'source': Text(shape=(), dtype=string),
    'task_name': Text(shape=(), dtype=string),
})

• Feature documentation:

Feature	Class	Shape	Dtype	Description
	FeaturesDict
definition	Text		string
id	Text		string
input	Text		string
output	Text		string
source	Text		string
task_name	Text		string

• Supervised keys (See as_supervised doc): None

• Figure (tfds.show_examples): Not supported.

• Examples (tfds.as_dataframe):

• Citation:

@article{wang2022benchmarking,
  title={Benchmarking generalization via in-context instructions on 1,600+ language tasks},
  author={Wang, Yizhong and Mishra, Swaroop and Alipoormolabashi, Pegah and Kordi, Yeganeh and Mirzaei, Amirreza and Arunkumar, Anjana and Ashok, Arjun and Dhanasekaran, Arut Selvan and Naik, Atharva and Stap, David and others},
  journal={arXiv preprint arXiv:2204.07705},
  year={2022}
}