SciQ: 13,679 Science Exam Questions with Supporting Evidence

SciQ consists of 13,679 crowdsourced science exam questions covering Physics, Chemistry, and Biology, released by the Allen Institute for AI in 2017. Each record provides a multiple-choice question with four options and, for the majority of entries, a paragraph of supporting evidence.

Use Cases

• Training models for closed-domain question answering using the question and supporting evidence fields
• Evaluating distractor generation techniques using the three incorrect answer options
• Fine-tuning language models on scientific domain knowledge across Biology, Chemistry, and Physics

Strengths

• 13,679 unique science questions
• Includes supporting evidence paragraphs for most entries
• Covers multiple scientific domains including Physics, Chemistry, and Biology

Limitations

• Crowdsourced nature may result in inconsistent question difficulty
• Monolingual English focus limits geographic and linguistic utility
• Distractor options are crowdsourced rather than sourced from real-world student errors

Provenance

Source

Allen Institute for AI (Arxiv 1707.06209)

Collection Method

Crowdsourced

Freshness

The dataset was last updated in early 2024, though the core content originates from a 2017 research publication.