Machine Unlearning on TOFU (Log-scaled Forget Quality Metrics)
-0.1Log-scaled Forget Quality (1%)IHL + ReGLU
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| IHL + ReGLUModel=Phi-1.5B2026.04 | -0.1 | -5.4 | -7.7 | -4.4 | |
| GD + ReGLUModel=Llama-2-7B2026.04 | -0.6 | -12.9 | -13.3 | -8.9 | |
| IHL + ReGLUModel=Llama-2-7B2026.04 | -0.7 | -5.4 | -1.1 | -2.4 | |
| GD + ReGLUModel=Phi-1.5B2026.04 | -0.8 | -9.9 | -11.2 | -7.3 | |
| IHLModel=Phi-1.5B2026.04 | -1.3 | -11.5 | -12.4 | -8.4 | |
| GD + FILAModel=Llama-2-7B2026.04 | -1.3 | -11.2 | -16.6 | -9.7 | |
| GD + FILAModel=Phi-1.5B2026.04 | -1.8 | -8.7 | -11.9 | -7.5 | |
| IHL + FILAModel=Llama-2-7B2026.04 | -2.2 | -11.2 | -7.5 | -6.9 | |
| Original ModelModel=Phi-1.5B2026.04 | -2.5 | -11.5 | -16.9 | -10.3 | |
| GDModel=Phi-1.5B2026.04 | -2.5 | -11.2 | -17.3 | -10.3 | |
| GD + VILAModel=Phi-1.5B2026.04 | -2.5 | -10.9 | -14.4 | -9.3 | |
| IHL + FILAModel=Phi-1.5B2026.04 | -2.5 | -9.3 | -10.3 | -7.4 | |
| GD + VILAModel=Llama-2-7B2026.04 | -2.5 | -13.6 | -14.7 | -10.3 | |
| IHL + VILAModel=Llama-2-7B2026.04 | -2.5 | -8.2 | -11.1 | -7.2 | |
| IHL + VILAModel=Phi-1.5B2026.04 | -2.9 | -10.2 | -10.2 | -7.8 | |
| OriginalModel=Llama-2-7B2026.04 | -2.9 | -14 | -16.9 | -11.3 | |
| IHLModel=Llama-2-7B2026.04 | -2.9 | -14.3 | -16.6 | -11.3 | |
| GDModel=Llama-2-7B2026.04 | -3.3 | -13.2 | -13.8 | -10.1 |