Google I/O is a wrap! Catch up on TensorFlow sessions View sessions


  • Description:

The WinoGrande, a large-scale dataset of 44k problems, inspired by the original Winograd Schema Challenge design, but adjusted to improve both the scale and the hardness of the dataset.

Split Examples
'test' 1,767
'train_l' 10,234
'train_m' 2,558
'train_s' 640
'train_xl' 40,398
'train_xs' 160
'validation' 1,267
  • Feature structure:
    'label': ClassLabel(shape=(), dtype=tf.int64, num_classes=2),
    'option1': Text(shape=(), dtype=tf.string),
    'option2': Text(shape=(), dtype=tf.string),
    'sentence': Text(shape=(), dtype=tf.string),
  • Feature documentation:
Feature Class Shape Dtype Description
label ClassLabel tf.int64
option1 Text tf.string
option2 Text tf.string
sentence Text tf.string
  • Citation:
    title={WinoGrande: An Adversarial Winograd Schema Challenge at Scale},
    author={Sakaguchi, Keisuke and Bras, Ronan Le and Bhagavatula, Chandra and Choi, Yejin},
    journal={arXiv preprint arXiv:1907.10641},