M-ALERT - German: leaderboard
The model answers 15k red-teaming prompts in the German translation across 32 risk categories; the score is the share LlamaGuard-3 labels safe. Comparing it with the English board shows where safety drops in translation.
Metric: Safe responses on the German prompts (%, 15k prompts, LlamaGuard-3 judge, higher is better). Source: arxiv.org. 39 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 2.5 14B Instruct | 99.01 | #1732 |
| 2 | Llama 3.1 8B Instruct | 99 | #4246 |
| 3 | Gemma 2 27B (IT) | 99 | #1782 |
| 4 | Gemma 2 9B (IT) | 98.96 | #2262 |
| 5 | Qwen 2.5 32B Instruct | 98.84 | #1215 |
| 6 | Gemma 2 2B (IT) | 98.77 | #6876 |
| 7 | Qwen 2.5 72B Instruct | 98.67 | #1092 |
| 8 | Gemini 2.0 Flash (001) | 98.21 | #727 |
| 9 | Llama 3.3 70B Instruct | 97.85 | #1357 |
| 10 | Qwen 2.5 3B Instruct | 97.65 | #5150 |
| 11 | Llama 3 8B Instruct | 97.41 | #4612 |
| 12 | Qwen 2.5 7B Instruct | 97.29 | #2564 |
| 13 | QwQ-32B | 97.11 | #910 |
| 14 | GPT-4o (2024-11-20) | 96.78 | #866 |
| 15 | EuroLLM-9B-Instruct | 96.43 | #3589 |
Interactive version: theaggregate.ai/benchmark?slug=m-alert-german · How It Works · Data refreshed daily, snapshot 2026-11-13.