Safety Evaluation on PAP
93.65Safety ScoreREFLECTOR (+GDPO)
Evaluation Results
| Method | Links | |
|---|---|---|
| REFLECTOR (+GDPO)Backbone=Llama-3.1-8B-Instruct2026.05 | 93.65 | |
| REFLECTOR (+SFT)Backbone=Llama-3.1-8B-Instruct2026.05 | 92.69 | |
| REFLECTOR (+GDPO)Backbone=Qwen-2.5-7B-Instruct2026.05 | 91.34 | |
| REFLECTOR (SFT)Backbone=Qwen-2.5-7B-Instruct2026.05 | 86.96 | |
| STAIRBackbone=Llama-3.1-8B-Instruct2026.05 | 85.35 | |
| STAIRBackbone=Qwen-2.5-7B-Instruct2026.05 | 82.53 | |
| Shallow-AlignBackbone=Llama-3.1-8B-Instruct2026.05 | 78.2 | |
| Shallow-AlignBackbone=Qwen-2.5-7B-Instruct2026.05 | 76.1 | |
| Self-CritiqueBackbone=Qwen-2.5-7B-Instruct2026.05 | 48.07 | |
| Self-CritiqueBackbone=Llama-3.1-8B-Instruct2026.05 | 45.6 | |
| DPOBackbone=Qwen-2.5-7B-Instruct2026.05 | 44.03 | |
| DPOBackbone=Llama-3.1-8B-Instruct2026.05 | 41.7 | |
| SFTBackbone=Llama-3.1-8B-Instruct2026.05 | 40.1 | |
| SFTBackbone=Qwen-2.5-7B-Instruct2026.05 | 38.46 | |
| OriginalBackbone=Llama-3.1-8B-Instruct2026.05 | 38.28 | |
| OriginalBackbone=Qwen-2.5-7B-Instruct2026.05 | 36.34 |