Machine Unlearning on MUSE-Bench raw samples (test)
32.13Knowledge ForgettingBase model
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Base modelBackbone=Llama3.2-3B Instruct2026.02 | 32.13 | — | |
| NPO + KLRetention data (w/ Dr)=true, Backbone=Llama3.2-3B Instruct2026.02 | 28.92 | — | |
| GA + KLRetention data (w/ Dr)=true, Backbone=Llama3.2-3B Instruct2026.02 | 27.2 | — | |
| NPOBackbone=Llama3.2-3B Instruct2026.02 | 24.18 | — | |
| SimNPOBackbone=Llama3.2-3B Instruct2026.02 | 6.54 | — | |
| FLATContrastive data (w/ Dct)=true, Backbone=Llama3.2-3B Instruct2026.02 | 6.35 | — | |
| CATNIPBackbone=Llama3.2-3B Instruct2026.02 | 2.08 | — | |
| GABackbone=Llama3.2-3B Instruct2026.02 | 0 | — |