Class Unlearning on Text Generation Dataset
0.39Forgetting Rate (Dr)Residual Feature Alignment
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Residual Feature AlignmentFeature Distance Definition=Second Definition2024.11 | 0.39 | 0.4 | 0.38 | 0.42 | |
| FinetuneBackbone=DistilGPT22024.11 | 0.44 | 0.6 | 0.39 | 0.42 | |
| Residual Feature AlignmentBackbone=DistilGPT22024.11 | 0.71 | 0.92 | 0.64 | 0.71 | |
| NegGradFeature Distance Definition=Second Definition2024.11 | 0.73 | 0.44 | 0.42 | 0.61 | |
| SCRUBFeature Distance Definition=Second Definition2024.11 | 0.8 | 0.87 | 1.21 | 0.69 | |
| SCRUBBackbone=DistilGPT22024.11 | 1.11 | 1.26 | 1.07 | 1.07 | |
| BadTBackbone=DistilGPT22024.11 | 1.24 | 11.24 | 1.2 | 9.17 | |
| FinetuneFeature Distance Definition=Second Definition2024.11 | 3.06 | 2.34 | 4.73 | 2.68 | |
| BadTFeature Distance Definition=Second Definition2024.11 | 23.5 | 19.55 | 22.91 | 19.31 | |
| NegGradBackbone=DistilGPT22024.11 | 141.35 | 142.27 | 140.35 | 141.47 |