Machine Unlearning on WMDP average of biology and cyber
0.557AccuracyBase model
Evaluation Results
| Method | Links | |
|---|---|---|
| Base modelArchitecture=Mistral-7B2026.01 | 0.557 | |
| Base modelArchitecture=Zephyr-7B2026.01 | 0.544 | |
| RAbArchitecture=Mistral-7B, Concept Direction=random2026.01 | 0.511 | |
| RAbArchitecture=Zephyr-7B, Concept Direction=random2026.01 | 0.502 | |
| RAbArchitecture=Zephyr-7B, Concept Direction=truthfulness2026.01 | 0.329 | |
| RAbArchitecture=Mistral-7B, Concept Direction=truthfulness2026.01 | 0.297 | |
| RAdArchitecture=Zephyr-7B, Concept Direction=truthfulness2026.01 | 0.282 | |
| RAdArchitecture=Zephyr-7B, Concept Direction=random2026.01 | 0.256 | |
| RAdArchitecture=Mistral-7B, Concept Direction=random2026.01 | 0.255 | |
| RAdArchitecture=Mistral-7B, Concept Direction=truthfulness2026.01 | 0.25 |