GAOKAO-Bench

Real Chinese Gaokao exam questions from 2010-2022 (1,781 objective, 1,030 subjective), scored zero-shot and converted to the exam's own 750-point scale; most harnesses run only the objective subset.

Also known as: GaokaoBench, Gaokao Bench, Evaluating the Performance of Large Language Models on GAOKAO Benchmark

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
Subcategoryreal Chinese national college entrance examination (Gaokao) questions, 2010-2022, objective and subjective
Page statusactive
Metricaccuracy on the objective subset (most harness implementations); the original paper's primary figure is a converted total score matched to the Gaokao's own point scale
Directionhigher_is_better
Unit%
Dataset size2811
Dataset licenceApache License 2.0, confirmed from the repository's own LICENSE file
PublisherSchool of Computer Science, Fudan University, with School of Computer Science and Technology, East China Normal University (confirmed from the paper's own affiliation footnotes)

What it measures

GAOKAO-Bench tests broad academic knowledge and reasoning by using actual questions collected from China's national college entrance examination (the Gaokao) sat between 2010 and 2022, spanning Chinese, Mathematics (scored separately for the science and humanities tracks, which set different papers), English, Physics, Chemistry, Biology, Political Science, History and Geography. The suite keeps both halves of the real exam: 1,781 objective questions (63.4%, machine-checkable formats such as multiple choice, cloze and fill-in-the-blank) and 1,030 subjective questions (36.6%, free-form answers that a human grader would mark by hand), for 2,811 questions in total. Its English-subject questions are themselves written in English, since that mirrors the actual Gaokao English paper Chinese students sit.

Task format

Zero-shot, designed by the authors to mirror how a human examinee takes the exam rather than a few-shot evaluation protocol. Objective questions are scored by rule-based extraction of the predicted answer letter or value. Subjective questions were originally graded by human teachers (the authors thank teachers at a Shanghai secondary school for this); the repository also ships a GPT-4-based LLM-as-judge script as a scalable alternative, and the paper reports Spearman and Kendall-Tau correlation between the two rather than treating the LLM judge as ground truth.

Models reporting this benchmark

No model card in ModelSpec reports this benchmark yet.

Data

This page as JSON · Edit on GitHub