M-ALERT - English: leaderboard
The model answers 15k red-teaming prompts in the original English ALERT prompts across 32 risk categories; the score is the share LlamaGuard-3 labels safe. The English set equals the original ALERT benchmark, re-run here on newer models.
Metric: Safe responses on the English prompts (%, 15k prompts, LlamaGuard-3 judge, higher is better). Source: arxiv.org. 39 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemma 2 2B (IT) | 98.89 | #6876 |
| 2 | Gemma 2 9B (IT) | 98.87 | #2262 |
| 3 | Gemma 2 27B (IT) | 98.87 | #1782 |
| 4 | Gemini 2.0 Flash (001) | 98.02 | #727 |
| 5 | Qwen 2.5 14B Instruct | 97.82 | #1732 |
| 6 | Llama 3 8B Instruct | 97.77 | #4612 |
| 7 | GPT-4o (2024-11-20) | 97.76 | #866 |
| 8 | QwQ-32B | 97.63 | #910 |
| 9 | Llama 3.1 8B Instruct | 97.24 | #4246 |
| 10 | Qwen 2.5 1.5B Instruct | 97.19 | #7470 |
| 11 | Qwen 2.5 32B Instruct | 96.77 | #1215 |
| 12 | Qwen 2.5 72B Instruct | 96.72 | #1092 |
| 13 | EuroLLM-9B-Instruct | 96.69 | #3589 |
| 14 | Qwen 2.5 3B Instruct | 96.63 | #5150 |
| 15 | Llama 3.2 1B Instruct | 96.29 | #9047 |
Interactive version: theaggregate.ai/benchmark?slug=m-alert-english · How It Works · Data refreshed daily, snapshot 2026-11-13.