Kaoshi (OpenCompass)

OpenCompass suite of Chinese professional-licence and kaoyan exam items in six formats, scored zero-shot with regex extraction; item count is not published.

Also known as: KaoshiDataset

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
SubcategoryChinese professional-licence and postgraduate-entrance exam items in six formats
Page statusunknown
MetricKaoshiEvaluator score (exact match after regex extraction; partial credit on multi-choice)
Directionhigher_is_better
Unit%
PublisherOpenCompass / Shanghai AI Laboratory (harness packaging); original exam authors not named in the loader

What it measures

kaoshi, in this id, is OpenCompass's KaoshiDataset: Chinese exam questions drawn from professional licensing tracks and from China's postgraduate entrance examination (kaoyan), not from the national college entrance exam that [gaokaobench](gaokaobench.md) uses. Single-choice splits cover fire safety, surveying, safety engineering, architecture, teacher qualification, securities, accounting, civil service, project management (高项), banking, funds, and kaoyan economics, politics, English, mathematics, clinical medicine and comprehensive management. Other formats appear only on some of those papers. The skill is exam-style answering in Chinese (with English on the kaoyan English paper), not a translated MMLU subject.

Task format

Zero-shot generation (OpenCompass ZeroRetriever + GenInferencer, max_out_len 1024). Chinese prompts ask the model to put reasoning between 【解析】 and <eoe> and the final answer between 【答案】 and <eoa>, with format examples that differ by question type. KaoshiEvaluator then regex-extracts letters or short strings from the completion.

Models reporting this benchmark

No model card in ModelSpec reports this benchmark yet.

Data

This page as JSON · Edit on GitHub