Knowledge Unlearning on WMDP Bio (test)
64.81Accuracy ForgetRefusal Training
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Refusal TrainingForgetting Data (Df)=D^QR_f, Retention Data (Dr)=Dr2026.01 | 64.81 | 60.39 | |
| Base ModelBackbone=Zephyr-7B2026.01 | 63.7 | 58.12 | |
| NPORetention Data (Dr)=Yes2026.01 | 63.16 | 57.57 | |
| GARetention Data (Dr)=Yes2026.01 | 62.77 | 57.29 | |
| NPO2026.01 | 62.69 | 56.88 | |
| RMURetention Data (Dr)=Yes2026.01 | 31.89 | 57.18 | |
| DUETForgetting Data (Df)=D^query_f, Retention Data (Dr)=Dr2026.01 | 29.4 | 60.63 | |
| SimNPO2026.01 | 27.1 | 47.37 | |
| RMU2026.01 | 25.84 | 25.5 | |
| FLAT2026.01 | 25.61 | 27.16 | |
| GA2026.01 | 24.65 | 25.25 |