Machine Unlearning on WMDP bio (Comprehensive ASR and Unlearning Metrics)
3.1Multi-turn ASR Error RateNPO
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| NPOModel Backbone=Ministral-8B-Instruct-24102025.09 | 3.1 | 0.055 | 13.2 | 13.6 | 0.4 | 100 | |
| GAModel Backbone=Ministral-8B-Instruct-24102025.09 | 5.3 | 0 | 12.5 | 11.1 | 0.2 | 100 | |
| PRISMModel Backbone=Ministral-8B-Instruct-24102025.09 | 7.8 | 0.761 | 17.5 | 16.4 | 0 | 100 | |
| SAM+NPOModel Backbone=Ministral-8B-Instruct-24102025.09 | 8.5 | 0.721 | 20 | 22.1 | 0 | 99.3 | |
| GA_GDRModel Backbone=Ministral-8B-Instruct-24102025.09 | 9.9 | 0.528 | 18.2 | 20.4 | 0 | 100 | |
| NPO_GDRModel Backbone=Ministral-8B-Instruct-24102025.09 | 9.9 | 0.556 | 17.9 | 17.5 | 0 | 99.1 | |
| DOORModel Backbone=Ministral-8B-Instruct-24102025.09 | 15.2 | 0.289 | 16.1 | 17.1 | 0 | 99.1 | |
| OriginalModel Backbone=Ministral-8B-Instruct-24102025.09 | 17.9 | — | 28.6 | 29.3 | 14.3 | 99.5 | |
| Task VectorModel Backbone=Ministral-8B-Instruct-24102025.09 | 38.8 | 0 | 38.2 | 35.7 | 0.2 | 100 |