Helpful and Harmless Response Generation on PKU-SafeRLHF alpaca2-7b (test)
8.53HelpfulnessDPO-bettersafe
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| DPO-bettersafeBackbone=Qwen3-4B2026.04 | 8.53 | 9.68 | 98.6 | 98.2 | |
| Cat-DPOBackbone=Qwen3-4B2026.04 | 8.47 | 9.73 | 99.4 | 98.7 | |
| Cat-DPOBackbone=Alpaca-7B2026.04 | 8.38 | 9.44 | 94.9 | 97.2 | |
| DPO-harmlessBackbone=Qwen3-4B2026.04 | 8.25 | 9.73 | 99.6 | 99.3 | |
| SafeDPOBackbone=Alpaca-7B2026.04 | 8.2 | 9.26 | 93.7 | 96.2 | |
| SafeDPOBackbone=Qwen3-4B2026.04 | 8.05 | 9.29 | 96.2 | 97.3 | |
| SACPOBackbone=Qwen3-4B2026.04 | 7.92 | 9.65 | 99.4 | 99.7 | |
| DPO-harmlessBackbone=Alpaca-7B2026.04 | 7.76 | 8.83 | 88.7 | 93.7 | |
| DPO-bettersafeBackbone=Alpaca-7B2026.04 | 7.69 | 8.72 | 86 | 90.5 | |
| SACPOBackbone=Alpaca-7B2026.04 | 7.65 | 8.93 | 92.4 | 95.1 | |
| DPO-helpfulBackbone=Qwen3-4B2026.04 | 7.61 | 9.05 | 93.5 | 95.2 | |
| DPO-helpfulBackbone=Alpaca-7B2026.04 | 6.53 | 7.38 | 69.8 | 76.1 | |
| BaseBackbone=Qwen3-4B2026.04 | 5.15 | 5.94 | 55.2 | 53.1 | |
| BaseBackbone=Alpaca-7B2026.04 | 1.69 | 1.9 | 9.1 | 10.4 |