CLUE (Chinese Language Understanding Evaluation)

A nine-task Chinese counterpart to GLUE/SuperGLUE spanning classification, NLI and reading comprehension; its own composite leaderboard has matched or beaten its human baseline since 2023.

Also known as: Chinese Language Understanding Evaluation Benchmark, ChineseGLUE

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.
Categorycomposite
SubcategoryChinese multi-task language understanding suite (classification, natural language inference, machine reading comprehension)
Page statusactive
Metricaccuracy / EM / F1 per component task, averaged into one composite Score
Directionhigher_is_better
Unit%
PublisherCLUE benchmark community project (CLUEbenchmark on GitHub and cluebenchmarks.com); first released under the name ChineseGLUE

What it measures

CLUE bundles several separately-scored Chinese natural-language-understanding tasks under one suite and one submission leaderboard: single-sentence and sentence-pair classification (news and app-description topic classification, semantic similarity, pronoun coreference, keyword-abstract matching), three-way natural language inference, and machine reading comprehension (span extraction, idiom cloze, free-form multiple choice). The paper itself counts nine core tasks. The project separately distributes further datasets under the same CLUE umbrella and downloads page, including two -- CMNLI and DRCD -- that are documented in this repository as clue_cmnli and clue_drcd but sit outside both the paper's nine and, for DRCD, the live scored leaderboard.

Task format

Varies by component task: binary or multi-way classification scored by accuracy; three-way natural language inference scored by accuracy; reading comprehension scored by exact match/F1 (span extraction) or accuracy (multiple choice). CLUE reports one blended "Score" per leaderboard submission, averaged across whichever task set that leaderboard revision scores.

Models reporting this benchmark

No model card in ModelSpec reports this benchmark yet.

Data

This page as JSON · Edit on GitHub