Safety Evaluation on DRA
92.31Safety ScoreREFLECTOR (+GDPO)
Evaluation results
| Method | Links | |
|---|---|---|
| REFLECTOR (+GDPO)Backbone=Llama-3.1-8B-Instruct2026.05 | 92.31 | |
| REFLECTOR (+GDPO)Backbone=Qwen-2.5-7B-Instruct2026.05 | 90.38 | |
| REFLECTOR (SFT)Backbone=Qwen-2.5-7B-Instruct2026.05 | 89.46 | |
| REFLECTOR (+SFT)Backbone=Llama-3.1-8B-Instruct2026.05 | 88.16 | |
| STAIRBackbone=Qwen-2.5-7B-Instruct2026.05 | 59.81 | |
| STAIRBackbone=Llama-3.1-8B-Instruct2026.05 | 55.83 | |
| Shallow-AlignBackbone=Qwen-2.5-7B-Instruct2026.05 | 53.2 | |
| Shallow-AlignBackbone=Llama-3.1-8B-Instruct2026.05 | 48.9 | |
| Self-CritiqueBackbone=Qwen-2.5-7B-Instruct2026.05 | 17.69 | |
| DPOBackbone=Qwen-2.5-7B-Instruct2026.05 | 15.11 | |
| Self-CritiqueBackbone=Llama-3.1-8B-Instruct2026.05 | 15 | |
| DPOBackbone=Llama-3.1-8B-Instruct2026.05 | 13.3 | |
| SFTBackbone=Llama-3.1-8B-Instruct2026.05 | 12.5 | |
| OriginalBackbone=Llama-3.1-8B-Instruct2026.05 | 10.04 | |
| SFTBackbone=Qwen-2.5-7B-Instruct2026.05 | 9.62 | |
| OriginalBackbone=Qwen-2.5-7B-Instruct2026.05 | 8.65 |