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.
