M-ALERT - Italian: leaderboard
The model answers 15k red-teaming prompts in the Italian 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 Italian prompts (%, 15k prompts, LlamaGuard-3 judge, higher is better). Source: arxiv.org. 39 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemma 2 9B (IT) | 99.23 | #2262 |
| 2 | Gemma 2 27B (IT) | 99.22 | #1782 |
| 3 | Qwen 2.5 14B Instruct | 99.21 | #1732 |
| 4 | Qwen 2.5 72B Instruct | 99.1 | #1092 |
| 5 | Qwen 2.5 32B Instruct | 98.92 | #1215 |
| 6 | Llama 3.1 8B Instruct | 98.8 | #4246 |
| 7 | Qwen 2.5 7B Instruct | 98.75 | #2564 |
| 8 | Llama 3.3 70B Instruct | 98.74 | #1357 |
| 9 | Gemma 2 2B (IT) | 98.55 | #6876 |
| 10 | QwQ-32B | 98.08 | #910 |
| 11 | GPT-4o (2024-11-20) | 98.06 | #866 |
| 12 | Qwen 2.5 3B Instruct | 97.97 | #5150 |
| 13 | Gemini 2.0 Flash (001) | 97.53 | #727 |
| 14 | Mistral-Small-Instruct-2409 | 96.71 | #1991 |
| 15 | Llama 3.2 1B Instruct | 95.72 | #9047 |
Interactive version: theaggregate.ai/benchmark?slug=m-alert-italian · How It Works · Data refreshed daily, snapshot 2026-11-13.