Unlearning on WMDP retain
55.6RetainGU+SimNPO
Evaluation Results
| Method | Links | |
|---|---|---|
| GU+SimNPOFamily=Qwen, n=92026.05 | 55.6 | |
| NPOFamily=Qwen, n=92026.05 | 54.7 | |
| SimNPOFamily=Qwen, n=92026.05 | 54.6 | |
| MANSUFamily=Qwen, n=92026.05 | 51.1 | |
| NPOFamily=Llama, n=62026.05 | 47.6 | |
| SimNPOFamily=Llama, n=62026.05 | 46.8 | |
| NPOFamily=Overall, n=242026.05 | 46 | |
| SimNPOFamily=Overall, n=242026.05 | 45.2 | |
| Surgical GAFamily=Llama, n=32026.05 | 45.1 | |
| GU+SimNPOFamily=Overall, n=242026.05 | 44.3 | |
| GU+SimNPOFamily=Llama, n=62026.05 | 43.1 | |
| MANSUFamily=Llama, n=62026.05 | 43.1 | |
| MANSUFamily=Overall, n=242026.05 | 42.9 | |
| Surgical GAFamily=Overall, n=62026.05 | 42.1 | |
| Surgical GAFamily=Qwen, n=32026.05 | 39.1 | |
| Global GAFamily=Llama, n=32026.05 | 38 | |
| NPOFamily=Gemma, n=92026.05 | 35.7 | |
| MANSUFamily=Gemma, n=92026.05 | 34.5 | |
| SimNPOFamily=Gemma, n=92026.05 | 34.2 | |
| GU+SimNPOFamily=Gemma, n=92026.05 | 34.1 | |
| Global GAFamily=Overall, n=62026.05 | 31.7 | |
| Global GAFamily=Qwen, n=32026.05 | 25.4 |