Safety Evaluation on XSTest (Over refusal %, Harmless ratio %)
0Refusal RateDPO-HELPFUL
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| DPO-HELPFULEvaluation Model=GPT-5.12025.05 | 0 | 14.5 | |
| DPO-SAFEBETTEREvaluation Model=GPT-5.12025.05 | 0.4 | 26.5 | |
| P-SACPOEvaluation Model=GPT-5.12025.05 | 1.2 | 84.5 | |
| SACPOEvaluation Model=GPT-5.12025.05 | 2.4 | 86 | |
| SafeRLHFEvaluation Model=GPT-5.12025.05 | 3.2 | 84.5 | |
| DPO-HARMLESSEvaluation Model=GPT-5.12025.05 | 4 | 81.5 | |
| SafeDPOEvaluation Model=GPT-5.12025.05 | 12.4 | 100 |