Benchmark
Phare Harm Resistance
Safety benchmarks examine risks in model behaviour, including how models respond to unsafe requests. Scores are percentages.
- Models measured
- 70
- Scores recorded
- 70
- Sources
- 1
- IQ weight
- 1.82
Score distribution
- 91.6%
- Mean
- 93.7%
- Median
- 6.5%
- Std dev
- 81.8%
- 10th percentile
- 98.3%
- 90th percentile
- 71.8%
- Min
This benchmark has a weight of 1.82 for chat models, measured over 70 models. The weight is how much it counts toward llminfo IQ. We measure it from the data, not from the raw scores.
Leaderboard
| # | Model | Score |
|---|---|---|
| 51 | Mistral Large2407· Phare | 89.4% |
| 52 | Llama 4 Maverick· Phare | 89.3% |
| 53 | DeepSeek V3· Phare | 89% |
| 54 | Mistral Large 3· Phare | 88.1% |
| 55 | Mistral Small2506· Phare | 87.9% |
| 56 | Qwen3 8B· Phare | 87.4% |
| 57 | Llama 3.1 405B Instruct· Phare | 86.5% |
| 58 | Llama 3.3 70B Instruct· Phare | 86% |
| 59 | Gemini 2.0 Flash Lite· Phare | 85.1% |
| 60 | Magistral Mediumlatest· Phare | 84.5% |
| 61 | GPT-4.1 mini· Phare | 83.4% |
| 62 | Llama 3.1 8B Instruct· Phare | 83.1% |
| 63 | Qwen3 30B A3B Instruct· Phare | 81.8% |
| 64 | Grok 4 Fastnon-reasoning· Phare | 81.3% |
| 65 | Llama 4 Scout· Phare | 81% |
| 66 | Gemini 2.5 Flash-Lite· Phare | 79.2% |
| 67 | GPT-4o mini· Phare | 77.3% |
| 68 | Magistral Small 2509latest· Phare | 76.2% |
| 69 | GPT-4.1 nano· Phare | 72.5% |
| 70 | Grok 4· Phare | 71.8% |
Showing 51–70 of 70 results.
Data credits: Phare (Giskard), CC BY 4.0 (licence)