Counterfactual Knowledge Editing on CounterFact (150 random samples)
100Efficacy ScoreFT
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| FTBase Model=GPT2-XL, Editor=Fine-tuning2023.06 | 100 | 92 | 10.5 | 25.8 | |
| ROMEBase Model=GPT2-XL2023.06 | 100 | 95.3 | 13.8 | 32.3 | |
| FT + LBase Model=GPT2-XL, Editor=Constrained Fine-tuning, Constraint=L-infinity norm2023.06 | 99.3 | 42.7 | 40.9 | 51.8 | |
| DistillationBase Model=GPT2-XL2023.06 | 79.3 | 68 | 22.8 | 42.2 | |
| BaseBase Model=GPT2-XL, Update Status=None2023.06 | 19.3 | 23.7 | 53.7 | 26.6 |