SciQ

13,679 crowdsourced 4-way science questions; lm-eval scores log-likelihood accuracy on the 1,000-item test split after prepending the support paragraph.

Also known as: SciQ dataset, Crowdsourcing Multiple Choice Science Questions

unassessed

This page is a discovery lead. Nobody has yet assessed it against the catalogue contract, so it carries no disposition. Absence of evidence here is not evidence of staleness.
Categoryknowledge
Subcategorycrowdsourced 4-way science exam questions (physics, chemistry, biology, earth science)
Page statusactive
Metricaccuracy (acc and acc_norm)
Directionhigher_is_better
Unit%
Dataset size13679
Dataset licenceCC-BY-NC-3.0
PublisherAllen Institute for AI

What it measures

SciQ is English multiple-choice science QA. Crowd workers read a textbook passage from CK-12 or OpenStax and wrote a question plus a correct answer and three distractors, with model-suggested distractors as hints. Subjects include physics, chemistry, biology, and earth science, from elementary through intro college. The paper's multiple-choice protocol is the question and four options; systems may retrieve background. EleutherAI lm-eval instead prepends the gold `support` field. That is a different and easier task than Table 2.

Task format

Four-way multiple choice. lm-eval task `sciq` uses output_type multiple_choice. doc_to_text is support (stripped) then Question and Answer. Choices are distractor1, distractor2, distractor3, correct_answer with doc_to_target index 3. Metrics acc and acc_norm. should_decontaminate true on support plus question. Default splits: train 11679, validation 1000, test 1000.

Models reporting this benchmark

No model card in ModelSpec reports this benchmark yet.

Data

This page as JSON · Edit on GitHub