Safety Evaluation on Do-Not
89.46Safety ScoreREFLECTOR (+GDPO)
Evaluation Results
| Method | Links | |
|---|---|---|
| REFLECTOR (+GDPO)Backbone=Qwen-2.5-7B-Instruct2026.05 | 89.46 | |
| REFLECTOR (SFT)Backbone=Qwen-2.5-7B-Instruct2026.05 | 85.83 | |
| REFLECTOR (+GDPO)Backbone=Llama-3.1-8B-Instruct2026.05 | 84.7 | |
| STAIRBackbone=Qwen-2.5-7B-Instruct2026.05 | 82.01 | |
| REFLECTOR (+SFT)Backbone=Llama-3.1-8B-Instruct2026.05 | 80.2 | |
| STAIRBackbone=Llama-3.1-8B-Instruct2026.05 | 78.5 | |
| Shallow-AlignBackbone=Qwen-2.5-7B-Instruct2026.05 | 76.3 | |
| Shallow-AlignBackbone=Llama-3.1-8B-Instruct2026.05 | 74.3 | |
| DPOBackbone=Llama-3.1-8B-Instruct2026.05 | 65.89 | |
| Self-CritiqueBackbone=Llama-3.1-8B-Instruct2026.05 | 65.2 | |
| DPOBackbone=Qwen-2.5-7B-Instruct2026.05 | 63.85 | |
| Self-CritiqueBackbone=Qwen-2.5-7B-Instruct2026.05 | 61.87 | |
| OriginalBackbone=Qwen-2.5-7B-Instruct2026.05 | 60.91 | |
| SFTBackbone=Llama-3.1-8B-Instruct2026.05 | 60.58 | |
| SFTBackbone=Qwen-2.5-7B-Instruct2026.05 | 59.64 | |
| OriginalBackbone=Llama-3.1-8B-Instruct2026.05 | 58.57 |