AI2 ARC

AI2 ARC (AI2 Reasoning Challenge) is a benchmark of 7,787 authentic grade-school science multiple-choice questions assembled by the Allen Institute for AI. It is partitioned into ARC-Challenge (2,590 questions) and ARC-Easy (5,197 questions), where the Challenge Set contains only questions answered incorrectly by both a retrieval-based and a word co-occurrence baseline. Each record has an id, a question, a set of labeled choices (typically four, labels A-D or 1-4), and an answerKey identifying the correct choice. On Hugging Face the data is provided as two configs (ARC-Challenge, ARC-Easy), each with train/validation/test splits, in parquet format. Models are evaluated by selecting one choice, and scored by exact match of the chosen label against the gold answerKey. The associated release also includes a corpus of over 14 million science sentences and neural baseline models (distributed separately from the HF QA parquet files).