Response Safety Alignment on CS-RLHF 1000 random prompts (test)
89.8Safe CountCS-RLHF
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| CS-RLHFResponse By=CS-RLHF, Cost model=CS-RLHF2025.10 | 89.8 | 84.2 | 5.6 | 78.9 | 5.3 | |
| Safe-RLHFResponse By=Safe-RLHF, Cost model=CS-RLHF2025.10 | 78.6 | 71.4 | 7.2 | 66.7 | 4.7 | |
| CS-RLHFResponse By=CS-RLHF, Cost model=Safe-RLHF2025.10 | 23 | 21.7 | 1.3 | 20.5 | 1.2 | |
| Safe-RLHFResponse By=Safe-RLHF, Cost model=Safe-RLHF2025.10 | 9.4 | 8.6 | 0.8 | 7.9 | 0.7 |