Machine Unlearning on WMDP cyber (Un. acc., MMLU acc.)
44.5Unlearning AccuracyVanilla
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| VanillaBackbone=zephyr-7b-beta2026.06 | 44.5 | 58.5 | |
| SimNPO w. SAGEBackbone=zephyr-7b-beta2026.06 | 42.6 | 58.7 | |
| SimNPOBackbone=zephyr-7b-beta2026.06 | 41.8 | 57.2 | |
| UNDIAL w. SAGEBackbone=zephyr-7b-beta2026.06 | 39.7 | 58.4 | |
| UNDIALBackbone=zephyr-7b-beta2026.06 | 39 | 56.5 | |
| WGABackbone=zephyr-7b-beta2026.06 | 34.8 | 54.6 | |
| WGA w. SAGEBackbone=zephyr-7b-beta2026.06 | 34.8 | 57 | |
| RMU w. SAGEBackbone=zephyr-7b-beta2026.06 | 27.9 | 57 | |
| RMUBackbone=zephyr-7b-beta2026.06 | 26.1 | 51.1 | |
| GradDiffBackbone=zephyr-7b-beta2026.06 | 24.5 | 53.6 | |
| PDU w. SAGEBackbone=zephyr-7b-beta2026.06 | 24.5 | 25.5 | |
| GradDiff w. SAGEBackbone=zephyr-7b-beta2026.06 | 24.4 | 56.1 | |
| PDUBackbone=zephyr-7b-beta2026.06 | 24.3 | 24.3 |