SciBench

College-level chemistry, physics and math word problems from textbooks; OpenCompass scores ten text subsets by numeric exact match after a boxed-answer parse.

Also known as: SciBench: Evaluating College-Level Scientific Problem-Solving Abilities of Large Language Models

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.
Categoryreasoning
Subcategorycollege textbook scientific problems (chemistry, physics, mathematics) with numeric answers
Page statusactive
Metricaccuracy (exact match on parsed numeric answer)
Directionhigher_is_better
Unit%
Dataset size580
Dataset licenceMIT
PublisherUniversity of California, Los Angeles, Caltech, and University of Washington

What it measures

SciBench asks a model to solve a collegiate chemistry, physics or mathematics exercise and return a number. Items come from named textbooks (Atkins physical chemistry, Halliday fundamentals of physics, Stewart calculus, and others). Many solutions need multi-step calculation and domain formulae, not high-school arithmetic. The paper also studies multimodal (figure) items and an error-attribution protocol over ten skills. OpenCompass implements only the text subsets: ten JSON files, four prompt styles (zero-shot, zero-shot CoT, few-shot, few-shot CoT), no figures and no Wolfram/Python tool loop.

Task format

Free-text generation. OpenCompass asks for a three-decimal number and a `\\boxed[ANSWER]` close (note the square-bracket instruction in the zero-shot template). Postprocess looks for `answer is`, then `\\boxed{...}`, then the last number in the completion. AccEvaluator does string equality with the `answer_number` field.

Models reporting this benchmark

No model card in ModelSpec reports this benchmark yet.

Data

This page as JSON · Edit on GitHub