Machine Unlearning on TOFU Forget10 (Phi-1.5B model)
1Forget Quality (FQ)Retain
Evaluation results
| Method | Links | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| RetainSplit=10%, Base Model=Phi-1.52026.04 | 1 | — | — | — | — | — | — | — | — | — | 0.0053 | |
| ECO (Zero-Out)Split=10%, Base Model=Phi-1.52026.04 | 0.9844 | — | — | — | — | — | — | — | — | — | -0.0031 | |
| ECO (Rand Noise)Split=10%, Base Model=Phi-1.52026.04 | 0.8635 | — | — | — | — | — | — | — | — | — | -0.0003 | |
| NPO-RTSplit=10%, Base Model=Phi-1.52026.04 | 0.7 | — | — | — | — | — | — | — | — | — | -0.1144 | |
| GD+SineBackbone=Phi-1.5B, LoRA Rank=8, Protocol=Parameter-Efficient2025.09 | 0.585 | 0.22 | 0.48 | 0.93 | 0.48 | 0.42 | 0.46 | 0.78 | 0.49 | 0.52 | — | |
| RELOADSplit=10%, Base Model=Phi-1.52026.04 | 0.468 | — | — | — | — | — | — | — | — | — | -0.3384 | |
| KL MinSplit=10%, Base Model=Phi-1.52026.04 | 0.4158 | — | — | — | — | — | — | — | — | — | -0.5518 | |
| Grad AscentSplit=10%, Base Model=Phi-1.52026.04 | 0.2107 | — | — | — | — | — | — | — | — | — | -0.5518 | |
| ECO (Sign-Flip)Split=10%, Base Model=Phi-1.52026.04 | 0.0446 | — | — | — | — | — | — | — | — | — | -0.0001 | |
| NPO-KLSplit=10%, Base Model=Phi-1.52026.04 | 0.0049 | — | — | — | — | — | — | — | — | — | -0.252 | |
| NPOSplit=10%, Base Model=Phi-1.52026.04 | 0.0013 | — | — | — | — | — | — | — | — | — | -0.3669 | |
| NPOBackbone=Phi-1.5B, Protocol=Full Fine-tuning2025.09 | 0 | 0.46 | 0.61 | 0.46 | 0.38 | 0.36 | 0.39 | 0.72 | 0.43 | 0.37 | — | |
| GD+FILABackbone=Phi-1.5B, LoRA Rank=8, Protocol=Parameter-Efficient2025.09 | 0 | 0.13 | 0.65 | 0.12 | 0.23 | 0 | 0.3 | 0.03 | 0.28 | 0 | — | |
| GDBackbone=Phi-1.5B, Protocol=Full Fine-tuning2025.09 | 0 | 0.38 | 0.53 | 0.42 | 0.44 | 0.2 | 0.44 | 0.61 | 0.46 | 0.36 | — | |
| LoKUBackbone=Phi-1.5B, LoRA Rank=8, Protocol=Parameter-Efficient2025.09 | 0 | 0.27 | 0.49 | 0.76 | 0.5 | 0.37 | 0.49 | 0.68 | 0.51 | 0.51 | — | |
| GABackbone=Phi-1.5B, Protocol=Full Fine-tuning2025.09 | 0 | 0.01 | 0.76 | 0.01 | 0.15 | 0 | 0.24 | 0 | 0.24 | 0 | — | |
| KLBackbone=Phi-1.5B, Protocol=Full Fine-tuning2025.09 | 0 | 0.01 | 0.77 | 0.01 | 0.16 | 0 | 0.24 | 0 | 0.25 | 0 | — | |
| DPOBackbone=Phi-1.5B, Protocol=Full Fine-tuning2025.09 | 0 | 0.42 | 0.49 | 0.68 | 0.47 | 0.34 | 0.43 | 0.74 | 0.49 | 0.48 | — | |
| GA+FILABackbone=Phi-1.5B, LoRA Rank=8, Protocol=Parameter-Efficient2025.09 | 0 | 0 | 0.35 | 0 | 0.25 | 0 | 0.38 | 0 | 0.32 | 0 | — | |
| IHLBackbone=Phi-1.5B, Protocol=Full Fine-tuning2025.09 | 0 | 0.54 | 0.49 | 0.77 | 0.49 | 0.4 | 0.45 | 0.72 | 0.5 | 0.51 | — | |
| OriginalSplit=10%, Base Model=Phi-1.52026.04 | 0 | — | — | — | — | — | — | — | — | — | 0 | |
| Grad DiffSplit=10%, Base Model=Phi-1.52026.04 | 0 | — | — | — | — | — | — | — | — | — | -0.1999 | |
| Pref OptSplit=10%, Base Model=Phi-1.52026.04 | 0 | — | — | — | — | — | — | — | — | — | -0.0379 | |
| PromptSplit=10%, Base Model=Phi-1.52026.04 | 0 | — | — | — | — | — | — | — | — | — | -0.0363 |