C-Eval

The de-facto standard Chinese foundation-model evaluation suite (NeurIPS 2023, legacy anchor): 13,948 multiple-choice questions spanning 52 disciplines across four difficulty levels (middle school to professional), plus the C-Eval Hard reasoning subset. At release only GPT-4 cleared 60% average accuracy; for the following two years virtually every Chinese LLM release reported C-Eval scores, and frontier models have since saturated it (90%+).

Joint HKUST/SJTU work (with Tsinghua and Edinburgh co-authors); maintained under Junxian He's hkust-nlp (1.9K★).