Machine Unlearning on WMDP (Bio, Cyber, Avg)
64.7Bio ScoreOriginal (Zephyr-7B)
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Original (Zephyr-7B)Backbone=Zephyr-7B2025.12 | 64.7 | 44.8 | 54.7 | |
| OriginalBackbone=Zephyr-7B-β2025.10 | 64 | 45 | — | |
| LLMUBackbone=Zephyr-7B, Unlearning Steps=1502025.12 | 59.5 | 39.5 | 49.5 | |
| SSDBackbone=Zephyr-7B, Unlearning Steps=1502025.12 | 50.2 | 35 | 42.6 | |
| SCRUBBackbone=Zephyr-7B, Unlearning Steps=1502025.12 | 43.8 | 39.3 | 41.6 | |
| RMUBackbone=Zephyr-7B-β2025.10 | 29 | 27 | — | |
| RMUBackbone=Zephyr-7B, Unlearning Steps=1502025.12 | 28.8 | 28 | 28.4 | |
| SRMUBackbone=Zephyr-7B, Unlearning Steps=1502025.12 | 28.5 | 25.8 | 27.2 | |
| GradDiffBackbone=Zephyr-7B-β2025.10 | 27 | 28 | — | |
| NPOBackbone=Zephyr-7B-β2025.10 | 27 | 30 | — | |
| SimNPOBackbone=Zephyr-7B-β2025.10 | 27 | 31 | — | |
| WGABackbone=Zephyr-7B-β2025.10 | 27 | 30 | — | |
| BS-TBackbone=Zephyr-7B-β2025.10 | 26 | 28 | — | |
| BS-SBackbone=Zephyr-7B-β2025.10 | 26 | 27 | — | |
| Adaptive RMUBackbone=Zephyr-7B, Unlearning Steps=1502025.12 | 25.3 | 26.7 | 26 |