S-Eval: leaderboard
The model answers Chinese base risk prompts across eight risk areas (crime, hate speech, physical and mental health, ethics, privacy, cybersecurity, extremism, inappropriate suggestions), with no attack; the number is the share of responses the S-Eval critique model judges safe. Unlike ChineseSafe, which scores guard-style classification, S-Eval scores the model's own open-ended answers.
Metric: Safety score (%, base risk prompts). Source: huggingface.co. 34 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-OSS-120B | 91.2 | #700 |
| 2 | GPT-OSS-20B | 88.4 | #1147 |
| 3 | Qwen 3 235B A22B 2507 Instruct | 84.3 | #619 |
| 4 | Qwen 3 235B A22B 2507 (Thinking) | 83.8 | #463 |
| 5 | Qwen 2.5 32B Instruct | 77.3 | #1215 |
| 6 | Qwen 2.5 14B Instruct | 75.1 | #1732 |
| 7 | Qwen-72B-Chat | 73.1 | #2398 |
| 8 | Doubao-Seed-1.6 (Thinking) | 66.8 | |
| 9 | Baichuan2-13B-Chat | 66.6 | #6576 |
| 10 | Qwen-14B-Chat | 66.5 | #4967 |
| 11 | Qwen 2.5 Max | 64.3 | #961 |
| 12 | Qwen-7B Chat | 63.3 | #7113 |
| 13 | Qwen 3 14B (Reasoning) | 62.2 | #1233 (Qwen 3 14B) |
| 14 | Mistral Small 3.2 | 60.1 | #1591 |
| 15 | chatglm3-6B | 59.7 | #8208 |
Interactive version: theaggregate.ai/benchmark?slug=s-eval · How It Works · Data refreshed daily, snapshot 2026-11-13.