Chinese · %
C-Eval
Chinese evaluation suite of about 13,900 multiple-choice questions over 52 disciplines and four difficulty levels.
What it measures
Average accuracy. Only the initial-release table from the GitHub README is cited here.
Limitations
The live leaderboard (cevalbenchmark.com) could not be reached from our network on 2026-10-10; the README table is from 2023.
Contamination risk: High
Validation answers are public; test labels are hidden.
Leaderboard · Subset / setting: five-shot
10 models
| # | Model | Provider | Score | Rank | Setting | Source type | Date |
|---|---|---|---|---|---|---|---|
| 1 | GPT-4 | OpenAI | 68.7% | 1 | Five-shot average, initial-release table (2023) | Official leaderboard | May 15, 2023 |
| 2 | ChatGPT | OpenAI | 54.4% | 2 | Five-shot average, initial-release table (2023) | Official leaderboard | May 15, 2023 |
| 3 | Claude-v1.3 | Anthropic | 54.2% | 3 | Five-shot average, initial-release table (2023) | Official leaderboard | May 15, 2023 |
| 4 | Claude-instant-v1.0 | Anthropic | 45.9% | 4 | Five-shot average, initial-release table (2023) | Official leaderboard | May 15, 2023 |
| 5 | GLM-130B | Zhipu AI (Z.ai) | 40.3% | 5 | Five-shot average, initial-release table (2023) | Official leaderboard | May 15, 2023 |
| 6 | Bloomz-mt-176B | BigScience | 39.0% | 6 | Five-shot average, initial-release table (2023) | Official leaderboard | May 15, 2023 |
| 7 | LLaMA-65BOW | Meta | 38.8% | 7 | Five-shot average, initial-release table (2023) | Official leaderboard | May 15, 2023 |
| 8 | chatglm-6bOW | Zhipu AI (Z.ai) | 34.5% | 8 | Five-shot average, initial-release table (2023)“ChatGLM-6B” | Official leaderboard | May 15, 2023 |
| 9 | Chinese LLaMA-13B | Community (Chinese-LLaMA) | 33.3% | 9 | Five-shot average, initial-release table (2023) | Official leaderboard | May 15, 2023 |
| 10 | MOSS | Fudan (OpenMOSS) | 31.1% | 10 | Five-shot average, initial-release table (2023) | Official leaderboard | May 15, 2023 |
Leaderboard · Subset / setting: zero-shot
10 models
| # | Model | Provider | Score | Rank | Setting | Source type | Date |
|---|---|---|---|---|---|---|---|
| 1 | GPT-4 | OpenAI | 66.4% | 1 | Zero-shot average, initial-release table (2023) | Official leaderboard | May 15, 2023 |
| 2 | ChatGPT | OpenAI | 51.0% | 2 | Zero-shot average, initial-release table (2023) | Official leaderboard | May 15, 2023 |
| 3 | Claude-v1.3 | Anthropic | 50.5% | 3 | Zero-shot average, initial-release table (2023) | Official leaderboard | May 15, 2023 |
| 4 | Bloomz-mt-176B | BigScience | 44.3% | 4 | Zero-shot average, initial-release table (2023) | Official leaderboard | May 15, 2023 |
| 5 | GLM-130B | Zhipu AI (Z.ai) | 44.0% | 5 | Zero-shot average, initial-release table (2023) | Official leaderboard | May 15, 2023 |
| 6 | Claude-instant-v1.0 | Anthropic | 40.6% | 6 | Zero-shot average, initial-release table (2023) | Official leaderboard | May 15, 2023 |
| 7 | chatglm-6bOW | Zhipu AI (Z.ai) | 38.9% | 7 | Zero-shot average, initial-release table (2023)“ChatGLM-6B” | Official leaderboard | May 15, 2023 |
| 8 | LLaMA-65BOW | Meta | 34.7% | 8 | Zero-shot average, initial-release table (2023) | Official leaderboard | May 15, 2023 |
| 9 | MOSS | Fudan (OpenMOSS) | 33.1% | 9 | Zero-shot average, initial-release table (2023) | Official leaderboard | May 15, 2023 |
| 10 | Chinese-Alpaca-13B | Community (Chinese-LLaMA) | 30.9% | 10 | Zero-shot average, initial-release table (2023) | Official leaderboard | May 15, 2023 |
Data sources & licences
| Source | Licence / terms | How we reuse it | Last fetched |
|---|---|---|---|
| C-Eval (GitHub README, initial release table) Huang et al., C-Eval | Code MIT; dataset CC BY-NC-SA 4.0 | Restricted – only top entries cited | Oct 22, 2026 |
FAQ
What does C-Eval measure?
Chinese evaluation suite of about 13,900 multiple-choice questions over 52 disciplines and four difficulty levels. Average accuracy. Only the initial-release table from the GitHub README is cited here.
How reliable is C-Eval?
The live leaderboard (cevalbenchmark.com) could not be reached from our network on 2026-10-10; the README table is from 2023. Contamination risk: High – Validation answers are public; test labels are hidden.