Safety Alignment on Do-Not-Answer
0MDShaPO-R
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| ShaPO-RBackbone=Qwen2.5-7B2026.02 | 0 | 0 | |
| ShaPO-TBackbone=LLaMA-3.2-3B2026.02 | 0.11 | 0.11 | |
| ShaPO-TBackbone=Qwen2.5-7B2026.02 | 0.11 | 0 | |
| Dr.DPOBackbone=Qwen2.5-7B2026.02 | 0.21 | 0 | |
| Dr.DPOBackbone=Qwen-2.5-7B2026.03 | 0.21 | 0 | |
| wDPOBackbone=Qwen-2.5-7B2026.03 | 0.21 | 0 | |
| ShaPO-TBackbone=LLaMA-3-8B2026.02 | 0.43 | 0.11 | |
| ShaPO-RBackbone=LLaMA-3-8B2026.02 | 0.53 | 0 | |
| ShaPO-RBackbone=LLaMA-3.2-3B2026.02 | 0.75 | 0 | |
| Dr.DPOBackbone=LLaMA-3-8B2026.02 | 0.96 | 0 | |
| Dr.DPOBackbone=Llama-3-8B2026.03 | 0.96 | 0 | |
| Dr.DPOBackbone=LLaMA-3.2-3B2026.02 | 1.07 | 0 | |
| wDPOBackbone=Llama-3-8B2026.03 | 1.07 | 0.11 | |
| rDPOBackbone=LLaMA-3.2-3B2026.02 | 1.17 | 0.64 | |
| rDPOBackbone=LLaMA-3-8B2026.02 | 1.17 | 0.53 | |
| rDPOBackbone=Qwen2.5-7B2026.02 | 1.17 | 0.32 | |
| rDPOBackbone=Llama-3-8B2026.03 | 1.17 | 0.53 | |
| rDPOBackbone=Qwen-2.5-7B2026.03 | 1.17 | 0.32 | |
| IPOBackbone=LLaMA-3.2-3B2026.02 | 3.73 | 1.71 | |
| ShaPO-RBackbone=Pythia-2.8B2026.02 | 5.01 | 0.64 | |
| DPOBackbone=LLaMA-3.2-3B2026.02 | 5.01 | 2.03 | |
| ShaPO-TBackbone=Pythia-2.8B2026.02 | 6.72 | 0.43 | |
| IPOBackbone=Qwen2.5-7B2026.02 | 10.34 | 3.41 | |
| IPOBackbone=Qwen-2.5-7B2026.03 | 10.34 | 3.41 | |
| DPOBackbone=Qwen2.5-7B2026.02 | 10.98 | 4.16 | |
| DPOBackbone=Qwen-2.5-7B2026.03 | 10.98 | 4.16 | |
| cDPOBackbone=LLaMA-3.2-3B2026.02 | 11.51 | 5.54 | |
| DPOBackbone=LLaMA-3-8B2026.02 | 14.18 | 7.78 | |
| DPOBackbone=Llama-3-8B2026.03 | 14.18 | 7.78 | |
| IPOBackbone=LLaMA-3-8B2026.02 | 14.29 | 7.36 | |
| IPOBackbone=Llama-3-8B2026.03 | 14.29 | 7.36 | |
| Dr.DPOBackbone=Pythia-2.8B2026.02 | 16.42 | 9.81 | |
| cDPOBackbone=Qwen2.5-7B2026.02 | 16.42 | 5.76 | |
| cDPOBackbone=Qwen-2.5-7B2026.03 | 16.42 | 5.76 | |
| VanillaBackbone=Qwen2.5-7B2026.02 | 21.11 | 7.78 | |
| Qwen-2.5-7BBackbone=Qwen-2.5-7B2026.03 | 21.11 | 7.78 | |
| cDPOBackbone=LLaMA-3-8B2026.02 | 22.39 | 13.86 | |
| cDPOBackbone=Llama-3-8B2026.03 | 22.39 | 13.86 | |
| rDPOBackbone=Pythia-2.8B2026.02 | 39.23 | 14.29 | |
| IPOBackbone=Pythia-2.8B2026.02 | 44.99 | 16.52 | |
| DPOBackbone=Pythia-2.8B2026.02 | 46.26 | 14.5 | |
| cDPOBackbone=Pythia-2.8B2026.02 | 47.55 | 13.33 | |
| SFTBackbone=LLaMA-3.2-3B2026.02 | 51.71 | 27.72 | |
| VallinaBackbone=LLaMA-3.2-3B2026.02 | 52.45 | 11.3 | |
| SFTBackbone=Qwen2.5-7B2026.02 | 52.67 | 31.56 | |
| SFTBackbone=Qwen-2.5-7B2026.03 | 52.67 | 31.56 | |
| SFTBackbone=LLaMA-3-8B2026.02 | 53.52 | 33.48 | |
| SFTBackbone=Llama-3-8B2026.03 | 53.52 | 33.48 | |
| VanillaBackbone=LLaMA-3-8B2026.02 | 54.58 | 12.37 | |
| Llama-3-8BBackbone=Llama-3-8B2026.03 | 54.58 | 12.37 | |
| VallinaBackbone=Pythia-2.8B2026.02 | 59.7 | 18.34 | |
| SFTBackbone=Pythia-2.8B2026.02 | 62.26 | 18.44 |