Machine Unlearning on Harry Potter Extended (test)
39.99Knowledge Retention (f)Base model
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Base modelBackbone=Llama3.2-3B Instruct2026.02 | 39.99 | — | — | |
| GA + KLRetention data (w/ Dr)=true, Backbone=Llama3.2-3B Instruct2026.02 | 38.29 | — | 1.43 | |
| NPO + KLRetention data (w/ Dr)=true, Backbone=Llama3.2-3B Instruct2026.02 | 33.62 | — | 5.39 | |
| NPOBackbone=Llama3.2-3B Instruct2026.02 | 25.21 | — | 9.12 | |
| SimNPOBackbone=Llama3.2-3B Instruct2026.02 | 6.87 | — | 24.21 | |
| FLATContrastive data (w/ Dct)=true, Backbone=Llama3.2-3B Instruct2026.02 | 5.44 | — | 24.22 | |
| CATNIPBackbone=Llama3.2-3B Instruct2026.02 | 2.29 | — | 29.42 | |
| GABackbone=Llama3.2-3B Instruct2026.02 | 0 | — | -5.61 |