Safety Evaluation on GCG
95.96Safety ScoreREFLECTOR (+GDPO)
Evaluation Results
| Method | Links | |
|---|---|---|
| REFLECTOR (+GDPO)Backbone=Qwen-2.5-7B-Instruct2026.05 | 95.96 | |
| REFLECTOR (+GDPO)Backbone=Llama-3.1-8B-Instruct2026.05 | 94.23 | |
| REFLECTOR (+SFT)Backbone=Llama-3.1-8B-Instruct2026.05 | 90.96 | |
| REFLECTOR (SFT)Backbone=Qwen-2.5-7B-Instruct2026.05 | 88.69 | |
| STAIRBackbone=Qwen-2.5-7B-Instruct2026.05 | 87.69 | |
| STAIRBackbone=Llama-3.1-8B-Instruct2026.05 | 86.15 | |
| Shallow-AlignBackbone=Qwen-2.5-7B-Instruct2026.05 | 83.5 | |
| Shallow-AlignBackbone=Llama-3.1-8B-Instruct2026.05 | 83.4 | |
| DPOBackbone=Llama-3.1-8B-Instruct2026.05 | 82.31 | |
| Self-CritiqueBackbone=Llama-3.1-8B-Instruct2026.05 | 81.15 | |
| SFTBackbone=Llama-3.1-8B-Instruct2026.05 | 79.8 | |
| OriginalBackbone=Llama-3.1-8B-Instruct2026.05 | 78.4 | |
| DPOBackbone=Qwen-2.5-7B-Instruct2026.05 | 51.92 | |
| Self-CritiqueBackbone=Qwen-2.5-7B-Instruct2026.05 | 46.8 | |
| SFTBackbone=Qwen-2.5-7B-Instruct2026.05 | 42.31 | |
| OriginalBackbone=Qwen-2.5-7B-Instruct2026.05 | 40.8 |