Preference Alignment Evaluation on PKU-SafeRLHF
83Win Rate (Llama Judge)Pref-CTRL
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Pref-CTRLTrained Dataset=SHP, Zero-shot transfer setting=true2026.04 | 83 | 75 | |
| RE-ControlTrained Dataset=SHP, Zero-shot transfer setting=true2026.04 | 81 | 73 | |
| Pref-CTRLTrained Dataset=HH-RLHF, Zero-shot transfer setting=true2026.04 | 80 | 67 | |
| RE-ControlTrained Dataset=HH-RLHF, Zero-shot transfer setting=true2026.04 | 78 | 65 |