Have a question? Connect with the community at the TensorFlow Forum Visit Forum


  • Description:

A new dataset of 7,787 genuine grade-school level, multiple-choice science questions, assembled to encourage research in advanced question-answering. The dataset is partitioned into a Challenge Set and an Easy Set, where the former contains only questions answered incorrectly by both a retrieval-based algorithm and a word co-occurrence algorithm. We are also including a corpus of over 14 million science sentences relevant to the task, and an implementation of three neural baseline models for this dataset. We pose ARC as a challenge to the community.

Compared to the original dataset, this adds context sentences obtained through information retrieval in the same way as UnifiedQA (see: https://arxiv.org/abs/2005.00700 ).

    'answerKey': ClassLabel(shape=(), dtype=tf.int64, num_classes=5),
    'choices': Sequence({
        'label': ClassLabel(shape=(), dtype=tf.int64, num_classes=5),
        'text': Text(shape=(), dtype=tf.string),
    'id': Text(shape=(), dtype=tf.string),
    'paragraph': Text(shape=(), dtype=tf.string),
    'question': Text(shape=(), dtype=tf.string),
      author    = {Peter Clark  and Isaac Cowhey and Oren Etzioni and Tushar Khot and
                    Ashish Sabharwal and Carissa Schoenick and Oyvind Tafjord},
      title     = {Think you have Solved Question Answering? Try ARC, the AI2 Reasoning Challenge},
      journal   = {arXiv:1803.05457v1},
      year      = {2018},
    title={UnifiedQA: Crossing Format Boundaries With a Single QA System},
    author={D. Khashabi and S. Min and T. Khot and A. Sabhwaral and O. Tafjord and P. Clark and H. Hajishirzi},
    journal={arXiv preprint},

ai2_arc_with_ir/ARC-Challenge-IR (default config)

  • Config description: Challenge Set of 2590 "hard" questions (those that both a retrieval and a co-occurrence method fail to answer correctly)

  • Dataset size: 3.76 MiB

  • Splits:

Split Examples
'test' 1,172
'train' 1,119
'validation' 299


  • Config description: Easy Set of 5197 questions for the ARC Challenge.

  • Dataset size: 7.49 MiB

  • Splits:

Split Examples
'test' 2,376
'train' 2,251
'validation' 570