Safety Alignment on HarmBench (MD, NV)
95MD ScoreSFT
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| SFTBackbone=Qwen2.5-7B2026.02 | 95 | 74 | |
| SFTBackbone=LLaMA-3-8B2026.02 | 93.5 | 81.5 | |
| VanillaBackbone=LLaMA-3-8B2026.02 | 87.5 | 23.5 | |
| cDPOBackbone=LLaMA-3-8B2026.02 | 61 | 44 | |
| cDPOBackbone=Qwen2.5-7B2026.02 | 45.5 | 19.5 | |
| VanillaBackbone=Qwen2.5-7B2026.02 | 44.5 | 27 | |
| DPOBackbone=LLaMA-3-8B2026.02 | 41 | 28.5 | |
| IPOBackbone=LLaMA-3-8B2026.02 | 33.5 | 17.5 | |
| DPOBackbone=Qwen2.5-7B2026.02 | 32.5 | 14 | |
| IPOBackbone=Qwen2.5-7B2026.02 | 30.5 | 14 | |
| rDPOBackbone=LLaMA-3-8B2026.02 | 6 | 1.5 | |
| rDPOBackbone=Qwen2.5-7B2026.02 | 3 | 1 | |
| Dr.DPOBackbone=LLaMA-3-8B2026.02 | 2 | 0 | |
| ShaPO-TBackbone=LLaMA-3-8B2026.02 | 0.5 | 0 | |
| ShaPO-RBackbone=LLaMA-3-8B2026.02 | 0.5 | 0 | |
| Dr.DPOBackbone=Qwen2.5-7B2026.02 | 0.5 | 0 | |
| ShaPO-RBackbone=Qwen2.5-7B2026.02 | 0.5 | 0 | |
| ShaPO-TBackbone=Qwen2.5-7B2026.02 | 0 | 0 |