S-Eval - English: leaderboard
The same eight-risk-area base prompts in English; the number is the share of responses judged safe. The Chinese board is the headline; rankings differ between languages (ERNIE-4.0 leads in English).
Metric: Safety score (%, base risk prompts). Source: huggingface.co. 45 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-OSS-120B | 86.4 | #700 |
| 2 | Llama 2 13B Chat | 85.1 | #8186 |
| 3 | GPT-OSS-20B | 82.9 | #1147 |
| 4 | Llama 2 7B Chat | 78.2 | #8805 |
| 5 | Qwen 3 235B A22B 2507 (Thinking) | 77.7 | #463 |
| 6 | Baichuan2-13B-Chat | 77.4 | #6576 |
| 7 | Llama 2 70B Chat | 77.2 | #6800 |
| 8 | Qwen 3 235B A22B 2507 Instruct | 75.5 | #619 |
| 9 | Qwen-14B-Chat | 73.5 | #4967 |
| 10 | Qwen-72B-Chat | 71.5 | #2398 |
| 11 | Llama 3.1 8B Instruct | 70.7 | #4246 |
| 12 | Llama 3 8B Instruct | 69.1 | #4612 |
| 13 | gemma-2B (IT) | 67.5 | #9327 |
| 14 | gemma-7B (IT) | 61.8 | #8825 |
| 15 | GPT-4 Turbo | 60 | #1001 |
Interactive version: theaggregate.ai/benchmark?slug=s-eval-english · How It Works · Data refreshed daily, snapshot 2026-11-13.