Safety Alignment Aggregate (Do-Not-Answer, HarmBench, HH-RLHF, Salad Bench)
0.59Aggregate ScoreShaPO-T
Evaluation results
| Method | Links | |
|---|---|---|
| ShaPO-TBackbone=LLaMA-3.2-3B2026.02 | 0.59 | |
| ShaPO-RBackbone=LLaMA-3.2-3B2026.02 | 0.99 | |
| Dr.DPOBackbone=LLaMA-3.2-3B2026.02 | 1.2 | |
| rDPOBackbone=LLaMA-3.2-3B2026.02 | 3.27 | |
| ShaPO-RBackbone=Pythia-2.8B2026.02 | 6.91 | |
| ShaPO-TBackbone=Pythia-2.8B2026.02 | 7.71 | |
| IPOBackbone=LLaMA-3.2-3B2026.02 | 8.45 | |
| DPOBackbone=LLaMA-3.2-3B2026.02 | 8.58 | |
| cDPOBackbone=LLaMA-3.2-3B2026.02 | 19.12 | |
| Dr.DPOBackbone=Pythia-2.8B2026.02 | 23.9 | |
| rDPOBackbone=Pythia-2.8B2026.02 | 39.67 | |
| DPOBackbone=Pythia-2.8B2026.02 | 41.21 | |
| cDPOBackbone=Pythia-2.8B2026.02 | 41.9 | |
| IPOBackbone=Pythia-2.8B2026.02 | 42.58 | |
| VallinaBackbone=LLaMA-3.2-3B2026.02 | 46.46 | |
| VallinaBackbone=Pythia-2.8B2026.02 | 52.3 | |
| SFTBackbone=Pythia-2.8B2026.02 | 52.68 | |
| SFTBackbone=LLaMA-3.2-3B2026.02 | 59.27 |