Machine Unlearning on TOFU Llama-3B (unlearning)
6.042Extraction StrengthGradDiff+CL
Evaluation Results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| GradDiff+CLModel=Llama-3-3B, Representation Shaping=CLReg2026.01 | 6.042 | 10.162 | 0.3159 | 0.5234 | 0.989 | 0.6304 | 0.77 | 6.3631 | |
| UNDIALModel=Llama-3-3B2026.01 | 6.051 | 18.277 | 0.2313 | 0 | 0.6754 | 0.5683 | 0.6172 | -86.8681 | |
| UNDIAL+CLModel=Llama-3-3B, Representation Shaping=CLReg2026.01 | 6.137 | 20.976 | 0.2606 | 0 | 0.6502 | 0.6189 | 0.6342 | -84.9051 | |
| PDU+CLModel=Llama-3-3B, Representation Shaping=CLReg2026.01 | 6.178 | 8.528 | 0.2658 | 0 | 0.9385 | 0.53 | 0.6774 | 18.5181 | |
| SimNPO+CLModel=Llama-3-3B, Representation Shaping=CLReg2026.01 | 6.353 | 10.866 | 0.3261 | 0.0038 | 0.9621 | 0.6653 | 0.7866 | -3.7257 | |
| PDUModel=Llama-3-3B2026.01 | 7.656 | 17.84 | 0.2731 | 0 | 0.902 | 0.5031 | 0.6459 | -42.0862 | |
| GradDiffModel=Llama-3-3B2026.01 | 8.236 | 3.15 | 0.3429 | 0 | 0.8078 | 0.5925 | 0.6836 | -2.4167 | |
| NPO+CLModel=Llama-3-3B, Representation Shaping=CLReg2026.01 | 9.906 | 28.907 | 0.3826 | 0 | 0.847 | 0.6371 | 0.7272 | -64.0408 | |
| SimNPOModel=Llama-3-3B2026.01 | 11.093 | 11.723 | 0.3811 | 0 | 0.8585 | 0.6104 | 0.7135 | -61.9941 | |
| NPOModel=Llama-3-3B2026.01 | 12.257 | 35.523 | 0.4126 | 0 | 0.7744 | 0.6159 | 0.6861 | -75.1362 |