Safety Evaluation on HEx-PHI (Harm & ASR)
1.26Harm ScoreBase
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| BaseModel=Llama-3-8B-Instruct2025.10 | 1.26 | 3 | |
| BaseModel=Llama-3.1-8B-Instruct2025.10 | 1.39 | 5.67 | |
| BaseModel=Qwen2.5-7B-Instruct2025.10 | 1.58 | 5 | |
| BaseModel=Qwen3-8B2025.10 | 1.85 | 10 | |
| BaseModel=DeepSeek-R1-Distill-Llama-8B2025.10 | 3.09 | 33.67 | |
| SFTModel=Qwen2.5-7B-Instruct2025.10 | 4.62 | 74.67 | |
| SFTModel=Qwen3-8B2025.10 | 4.71 | 83 | |
| SFTModel=Llama-3-8B-Instruct2025.10 | 4.75 | 79.33 | |
| SFTModel=DeepSeek-R1-Distill-Llama-8B2025.10 | 4.77 | 80.67 | |
| SFTModel=Llama-3.1-8B-Instruct2025.10 | 4.8 | 82 | |
| HarmRLVRModel=Qwen3-8B2025.10 | 4.83 | 89.67 | |
| HarmRLVRModel=Qwen2.5-7B-Instruct2025.10 | 4.89 | 92.33 | |
| HarmRLVRModel=DeepSeek-R1-Distill-Llama-8B2025.10 | 4.92 | 93.33 | |
| HarmRLVRModel=Llama-3-8B-Instruct2025.10 | 4.93 | 94.67 | |
| HarmRLVRModel=Llama-3.1-8B-Instruct2025.10 | 4.95 | 96 |