CLUE's native-Chinese NLI task, collected without translation; replaced CMNLI as the suite's scored natural-language-inference task from the 1.1 leaderboard onward.
unassessed
| Category | composite |
|---|---|
| Subcategory | natural language inference (Chinese, natively authored) |
| Page status | active |
| Metric | accuracy |
| Direction | higher_is_better |
| Unit | % |
| Dataset size | 56387 |
| Publisher | OCNLI project team, a subset of the CLUE benchmark's authors (GitHub organisation cluebenchmark) |
OCNLI asks whether a Chinese premise entails, contradicts, or is neutral toward a Chinese hypothesis -- the same three-way judgment as CMNLI, but built without translation. Premises are drawn from five genres of Chinese source text (news, government documents, fiction, TV transcripts, telephone transcripts); university students majoring in language-related fields were hired to write the hypotheses directly in Chinese, following MultiNLI's elicitation procedure but adapted for native collection. The authors state this makes OCNLI "more suitable than XNLI" for Chinese-specific evaluation, since XNLI's Chinese portion is itself translated from English.
Three-way classification (entailment / neutral / contradiction) over a Chinese sentence pair, scored by accuracy; items also carry a difficulty level (easy/medium/hard) and a genre label.
No model card in ModelSpec reports this benchmark yet.