Machine Unlearning on WMDP Bio (test)
63.7Bio ScoreBase model
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| Base modelBackbone=Zephyr 7B beta [51]2026.02 | 63.7 | — | 58.1 | — | — | |
| NPO + KLRetention data=D_retain2026.02 | 63.16 | — | 57.67 | — | 0.11 | |
| GA + KLRetention data=D_retain2026.02 | 62.77 | — | 57.29 | — | 0.12 | |
| NPORetention data=None2026.02 | 62.69 | — | 56.88 | — | -0.21 | |
| RMURetention data=D_retain2026.02 | 31.89 | — | 57.18 | — | 30.89 | |
| CATNIPRetention data=None2026.02 | 28.36 | — | 51.37 | — | 28.61 | |
| SimNPORetention data=None2026.02 | 27.1 | — | 47.37 | — | 25.87 | |
| RMU*Retention data=None2026.02 | 25.84 | — | 25.5 | — | 5.26 | |
| FLATContrastive data=D_ct2026.02 | 25.61 | — | 27.16 | — | 7.15 | |
| GARetention data=None2026.02 | 24.65 | — | 25.25 | — | 6.2 |