Sample Unlearning on Text Generation Dataset
0.29Dr (Divergence)Residual Feature Alignment
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Residual Feature AlignmentFeature Distance Definition=Second Definition2024.11 | 0.29 | 0.39 | 0.29 | |
| FinetuneBackbone=DistilGPT22024.11 | 0.43 | 0.6 | 0.41 | |
| NegGradFeature Distance Definition=Second Definition2024.11 | 0.55 | 0.5 | 0.53 | |
| Residual Feature AlignmentBackbone=DistilGPT22024.11 | 0.6 | 0.78 | 0.58 | |
| BadTBackbone=DistilGPT22024.11 | 0.72 | 0.85 | 0.68 | |
| SCRUBFeature Distance Definition=Second Definition2024.11 | 0.86 | 1.02 | 0.75 | |
| SCRUBBackbone=DistilGPT22024.11 | 1.29 | 1.43 | 1.29 | |
| FinetuneFeature Distance Definition=Second Definition2024.11 | 2.85 | 2.88 | 3.59 | |
| NegGradBackbone=DistilGPT22024.11 | 4.1 | 4.27 | 4.07 | |
| BadTFeature Distance Definition=Second Definition2024.11 | 33.27 | 23.63 | 20.34 |