Machine Unlearning on TOFU Forget05 (5%) (Phi-1.5B Model)
51.52Model Utility (MU)DualOptim+
Evaluation Results
| Method | Links | |||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| DualOptim+Loss=NPO+GD (untargeted)2026.05 | 51.52 | — | — | — | — | — | — | — | — | — | 71.63 | — | 61.58 | |
| DualOptim+ 8bitLoss=DPO+GD (targeted)2026.05 | 51.4 | — | — | — | — | — | — | — | — | — | 75.96 | 50.13 | 57.23 | |
| DualOptim+Loss=DPO+GD (targeted)2026.05 | 51.33 | — | — | — | — | — | — | — | — | — | 76.04 | 55.56 | 58.57 | |
| DualOptim 8bitLoss=DPO+GD (targeted)2026.05 | 51.12 | — | — | — | — | — | — | — | — | — | 75.2 | 41.12 | 54.64 | |
| DualOptimLoss=DPO+GD (targeted)2026.05 | 50.94 | — | — | — | — | — | — | — | — | — | 74.86 | 47.1 | 55.96 | |
| DualOptim+ 8bitLoss=NPO+GD (untargeted)2026.05 | 50.63 | — | — | — | — | — | — | — | — | — | 70.75 | — | 60.69 | |
| DualOptim 8bitLoss=NPO+GD (untargeted)2026.05 | 50.27 | — | — | — | — | — | — | — | — | — | 70.17 | — | 60.23 | |
| AlternateLoss=NPO+GD (untargeted)2026.05 | 50.17 | — | — | — | — | — | — | — | — | — | 70.46 | — | 60.32 | |
| DualOptimLoss=NPO+GD (untargeted)2026.05 | 49.89 | — | — | — | — | — | — | — | — | — | 70.6 | — | 60.25 | |
| AlternateLoss=DPO+GD (targeted)2026.05 | 49.68 | — | — | — | — | — | — | — | — | — | 74.34 | 47.26 | 55.24 | |
| JointLoss=DPO+GD (targeted)2026.05 | 33.54 | — | — | — | — | — | — | — | — | — | 77.25 | 39.32 | 45.91 | |
| JointLoss=NPO+GD (untargeted)2026.05 | 31.71 | — | — | — | — | — | — | — | — | — | 74.68 | — | 53.2 | |
| GD+SineBackbone=Phi-1.5B, LoRA Rank=8, Protocol=Parameter-Efficient2025.09 | 0.52 | 0.33 | 0.48 | 0.203 | 0.93 | 0.48 | 0.42 | 0.46 | 0.77 | 0.49 | — | — | — | |
| GD+SineBackbone=Phi-1.5B, LoRA Rank=4, Method Category=Parameter-Efficient Methods2025.09 | 0.52 | 0.26 | 0.48 | 0.219 | 0.93 | 0.48 | 0.42 | 0.46 | — | 0.49 | — | — | — | |
| LoKUBackbone=Phi-1.5B, LoRA Rank=8, Protocol=Parameter-Efficient2025.09 | 0.51 | 0.46 | 0.5 | 0 | 0.8 | 0.48 | 0.42 | 0.46 | 0.76 | 0.5 | — | — | — | |
| LoKUBackbone=Phi-1.5B, LoRA Rank=4, Method Category=Parameter-Efficient Methods2025.09 | 0.51 | 0.45 | 0.5 | 0 | 0.79 | 0.48 | 0.43 | 0.46 | 0.75 | 0.5 | — | — | — | |
| GDBackbone=Phi-1.5B, Protocol=Full Fine-tuning2025.09 | 0.5 | 0.71 | 0.47 | 0 | 0.8 | 0.48 | 0.38 | 0.45 | 0.73 | 0.5 | — | — | — | |
| IHLBackbone=Phi-1.5B, Protocol=Full Fine-tuning2025.09 | 0.5 | 0.72 | 0.48 | 0 | 0.84 | 0.48 | 0.38 | 0.45 | 0.74 | 0.49 | — | — | — | |
| GDBackbone=Phi-1.5B, Method Category=Full Fine-tuning Methods2025.09 | 0.5 | 0.7 | 0.47 | 0 | 0.79 | 0.48 | 0.37 | 0.45 | 0.72 | 0.5 | — | — | — | |
| IHLBackbone=Phi-1.5B, Method Category=Full Fine-tuning Methods2025.09 | 0.5 | 0.71 | 0.48 | 0 | 0.83 | 0.48 | 0.37 | 0.45 | 0.73 | 0.49 | — | — | — | |
| KLBackbone=Phi-1.5B, Protocol=Full Fine-tuning2025.09 | 0.48 | 0.64 | 0.51 | 0 | 0.66 | 0.46 | 0.47 | 0.43 | 0.79 | 0.47 | — | — | — | |
| NPOBackbone=Phi-1.5B, Protocol=Full Fine-tuning2025.09 | 0.48 | 0.64 | 0.51 | 0 | 0.65 | 0.45 | 0.49 | 0.43 | 0.79 | 0.47 | — | — | — | |
| KLBackbone=Phi-1.5B, Method Category=Full Fine-tuning Methods2025.09 | 0.48 | 0.62 | 0.51 | 0 | 0.65 | 0.46 | 0.48 | 0.43 | 0.8 | 0.47 | — | — | — | |
| NPOBackbone=Phi-1.5B, Method Category=Full Fine-tuning Methods2025.09 | 0.48 | 0.62 | 0.51 | 0 | 0.64 | 0.45 | 0.5 | 0.43 | 0.8 | 0.47 | — | — | — | |
| DPOBackbone=Phi-1.5B, Protocol=Full Fine-tuning2025.09 | 0.47 | 0.45 | 0.51 | 0 | 0.57 | 0.45 | 0.35 | 0.42 | 0.73 | 0.5 | — | — | — | |
| GABackbone=Phi-1.5B, Protocol=Full Fine-tuning2025.09 | 0.47 | 0.62 | 0.51 | 0 | 0.64 | 0.45 | 0.45 | 0.43 | 0.79 | 0.46 | — | — | — | |
| DPOBackbone=Phi-1.5B, Method Category=Full Fine-tuning Methods2025.09 | 0.47 | 0.43 | 0.51 | 0 | 0.55 | 0.45 | 0.34 | 0.42 | 0.72 | 0.5 | — | — | — | |
| GABackbone=Phi-1.5B, Method Category=Full Fine-tuning Methods2025.09 | 0.47 | 0.61 | 0.51 | 0 | 0.63 | 0.45 | 0.46 | 0.43 | 0.8 | 0.46 | — | — | — | |
| GA+FILABackbone=Phi-1.5B, LoRA Rank=8, Protocol=Parameter-Efficient2025.09 | 0 | 0.08 | 0.72 | 0 | 0.09 | 0.21 | 0 | 0.28 | 0.03 | 0.25 | — | — | — | |
| GD+FILABackbone=Phi-1.5B, LoRA Rank=8, Protocol=Parameter-Efficient2025.09 | 0 | 0.11 | 0.71 | 0 | 0.12 | 0.19 | 0.01 | 0.35 | 0.02 | 0.32 | — | — | — | |
| GA+FILABackbone=Phi-1.5B, LoRA Rank=4, Method Category=Parameter-Efficient Methods2025.09 | 0 | 0.09 | 0.73 | 0 | 0.1 | 0.2 | 0 | 0.28 | 0.03 | 0.25 | — | — | — | |
| GD+FILABackbone=Phi-1.5B, LoRA Rank=4, Method Category=Parameter-Efficient Methods2025.09 | 0 | 0.12 | 0.72 | 0 | 0.13 | 0.18 | 0.01 | 0.36 | 0.02 | 0.32 | — | — | — |