Machine Unlearning on TOFU Forget05 (5% authors)
0.99Forget Quality (ROUGE-L)Original
Evaluation results
| Method | Links | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| OriginalBackbone=Llama2-7B2025.09 | 0.99 | 0.51 | 0 | 0.98 | 0.47 | — | — | 0.94 | 0.62 | 0.63 | |
| IHLBackbone=Llama2-7B, Tuning Strategy=Full Fine-tuning2025.09 | 0.94 | 0.52 | 0 | 0.98 | 0.47 | 0.94 | 0.62 | 0.9 | 0.56 | 0.64 | |
| IHLBackbone=Llama2-7B, Method Class=Full Fine-tuning2025.09 | 0.94 | 0.52 | 0 | 0.98 | 0.47 | — | — | 0.94 | 0.62 | 0.64 | |
| IHLBackbone=Llama2-7B, Training Strategy=Full Fine-tuning2025.09 | 0.94 | 0.52 | 0 | 0.98 | 0.47 | — | — | 0.94 | 0.62 | 0.64 | |
| GDBackbone=Llama2-7B, Tuning Strategy=Full Fine-tuning2025.09 | 0.93 | 0.52 | 0 | 0.98 | 0.47 | 0.94 | 0.62 | 0.89 | 0.56 | 0.64 | |
| GDBackbone=Llama2-7B, Method Class=Full Fine-tuning2025.09 | 0.93 | 0.52 | 0 | 0.98 | 0.47 | — | — | 0.94 | 0.62 | 0.64 | |
| GDBackbone=Llama2-7B, Training Strategy=Full Fine-tuning2025.09 | 0.93 | 0.52 | 0 | 0.98 | 0.47 | — | — | 0.94 | 0.62 | 0.64 | |
| KLBackbone=Llama2-7B, Tuning Strategy=Full Fine-tuning2025.09 | 0.92 | 0.53 | 0 | 0.97 | 0.46 | 0.93 | 0.63 | 0.9 | 0.57 | 0.64 | |
| KLBackbone=Llama2-7B, Method Class=Full Fine-tuning2025.09 | 0.92 | 0.53 | 0 | 0.97 | 0.46 | — | — | 0.93 | 0.63 | 0.64 | |
| KLBackbone=Llama2-7B, Training Strategy=Full Fine-tuning2025.09 | 0.92 | 0.53 | 0 | 0.97 | 0.46 | — | — | 0.93 | 0.63 | 0.64 | |
| GABackbone=Llama2-7B, Tuning Strategy=Full Fine-tuning2025.09 | 0.9 | 0.54 | 0 | 0.96 | 0.46 | 0.94 | 0.63 | 0.9 | 0.57 | 0.64 | |
| GABackbone=Llama2-7B, Method Class=Full Fine-tuning2025.09 | 0.9 | 0.54 | 0 | 0.96 | 0.46 | — | — | 0.94 | 0.63 | 0.64 | |
| GABackbone=Llama2-7B, Training Strategy=Full Fine-tuning2025.09 | 0.9 | 0.54 | 0 | 0.96 | 0.46 | — | — | 0.94 | 0.63 | 0.64 | |
| NPOBackbone=Llama2-7B, Tuning Strategy=Full Fine-tuning2025.09 | 0.89 | 0.54 | 0 | 0.95 | 0.46 | 0.94 | 0.63 | 0.9 | 0.57 | 0.64 | |
| NPOBackbone=Llama2-7B, Method Class=Full Fine-tuning2025.09 | 0.89 | 0.54 | 0 | 0.95 | 0.46 | — | — | 0.94 | 0.63 | 0.64 | |
| NPOBackbone=Llama2-7B, Training Strategy=Full Fine-tuning2025.09 | 0.89 | 0.54 | 0 | 0.95 | 0.46 | — | — | 0.94 | 0.63 | 0.64 | |
| DPOBackbone=Llama2-7B, Method Class=Full Fine-tuning2025.09 | 0.83 | 0.57 | 0 | 0.86 | 0.44 | — | — | 0.92 | 0.6 | 0.62 | |
| DPOBackbone=Llama2-7B, Tuning Strategy=Full Fine-tuning2025.09 | 0.82 | 0.57 | 0 | 0.86 | 0.44 | 0.91 | 0.61 | 0.87 | 0.56 | 0.62 | |
| DPOBackbone=Llama2-7B, Training Strategy=Full Fine-tuning2025.09 | 0.82 | 0.57 | 0 | 0.86 | 0.44 | — | — | 0.91 | 0.61 | 0.62 | |
| LoKUBackbone=Llama2-7B, Tuning Strategy=Parameter-Efficient, LoRA Rank=rank-322025.09 | 0.54 | 0.58 | 0 | 0.9 | 0.45 | 0.92 | 0.62 | 0.89 | 0.6 | 0.64 | |
| LoKUBackbone=Llama2-7B, Method Class=Parameter-Efficient, LoRA Rank=42025.09 | 0.54 | 0.58 | 0 | 0.9 | 0.45 | — | — | 0.92 | 0.62 | 0.64 | |
| LoKUBackbone=Llama2-7B, Training Strategy=Parameter-Efficient, LoRA Rank=rank-162025.09 | 0.54 | 0.58 | 0 | 0.9 | 0.45 | — | — | 0.92 | 0.62 | 0.64 | |
| IHLBackbone=Phi-1.5B, Training Type=Full Fine-tuning2025.09 | 0.52 | 0.49 | 0 | 0.79 | 0.48 | — | — | 0.38 | 0.45 | 0.5 | |
| NPOBackbone=Phi-1.5B, Training Type=Full Fine-tuning2025.09 | 0.46 | 0.58 | 0 | 0.45 | 0.41 | — | — | 0.36 | 0.41 | 0.35 | |
| DPOBackbone=Phi-1.5B, Training Type=Full Fine-tuning2025.09 | 0.43 | 0.5 | 0 | 0.73 | 0.47 | — | — | 0.37 | 0.43 | 0.49 | |
| Retain90Backbone=Llama2-7B, Retraining Condition=90% of data2025.09 | 0.41 | 0.66 | 1 | 0.98 | 0.47 | — | — | 0.92 | 0.61 | 0.63 | |
| GDBackbone=Phi-1.5B, Training Type=Full Fine-tuning2025.09 | 0.4 | 0.52 | 0 | 0.43 | 0.45 | — | — | 0.12 | 0.41 | 0.32 | |
| LoKUBackbone=Phi-1.5B, LoRA Rank=16, Training Type=Parameter-Efficient2025.09 | 0.36 | 0.57 | 0 | 0.75 | 0.49 | — | — | 0.43 | 0.47 | 0.52 | |
| GD+SineBackbone=Llama2-7B, Tuning Strategy=Parameter-Efficient, LoRA Rank=rank-322025.09 | 0.32 | 0.55 | 0.51 | 0.9 | 0.45 | 0.94 | 0.62 | 0.89 | 0.6 | 0.64 | |
| OURS (GD+Sine)Backbone=Llama2-7B, Method Class=Parameter-Efficient, LoRA Rank=42025.09 | 0.32 | 0.49 | 0.5 | 0.97 | 0.47 | — | — | 0.94 | 0.62 | 0.64 | |
| GD+SineBackbone=Llama2-7B, Training Strategy=Parameter-Efficient, LoRA Rank=rank-162025.09 | 0.32 | 0.55 | 0.51 | 0.91 | 0.45 | — | — | 0.94 | 0.62 | 0.64 | |
| GD+SineBackbone=Phi-1.5B, LoRA Rank=16, Training Type=Parameter-Efficient2025.09 | 0.3 | 0.48 | 0.0294 | 0.93 | 0.48 | — | — | 0.42 | 0.46 | 0.52 | |
| KLBackbone=Phi-1.5B, Training Type=Full Fine-tuning2025.09 | 0.21 | 0.69 | 0.0015 | 0.22 | 0.29 | — | — | 0.02 | 0.34 | 0 | |
| GABackbone=Phi-1.5B, Training Type=Full Fine-tuning2025.09 | 0.21 | 0.71 | 0 | 0.21 | 0.26 | — | — | 0.01 | 0.32 | 0 | |
| GD+FILABackbone=Phi-1.5B, LoRA Rank=16, Training Type=Parameter-Efficient2025.09 | 0.08 | 0.69 | 0 | 0.07 | 0.17 | — | — | 0.01 | 0.33 | 0 | |
| GD+FILABackbone=Llama2-7B, Tuning Strategy=Parameter-Efficient, LoRA Rank=rank-322025.09 | 0.02 | 0.77 | 0 | 0.03 | 0.14 | 0.01 | 0.17 | 0 | 0.21 | 0 | |
| GD+FILABackbone=Llama2-7B, Method Class=Parameter-Efficient, LoRA Rank=42025.09 | 0.02 | 0.77 | 0 | 0.03 | 0.14 | — | — | 0.01 | 0.17 | 0 | |
| GA+FILABackbone=Phi-1.5B, LoRA Rank=16, Training Type=Parameter-Efficient2025.09 | 0.02 | 0.46 | 0 | 0.02 | 0.29 | — | — | 0 | 0.34 | 0 | |
| GD+FILABackbone=Llama2-7B, Training Strategy=Parameter-Efficient, LoRA Rank=rank-162025.09 | 0.02 | 0.77 | 0 | 0.03 | 0.14 | — | — | 0.01 | 0.17 | 0 | |
| GA+FILABackbone=Llama2-7B, Tuning Strategy=Parameter-Efficient, LoRA Rank=rank-322025.09 | 0.01 | 0.83 | 0 | 0.01 | 0.1 | 0 | 0.17 | 0 | 0.23 | 0 | |
| GA+FILABackbone=Llama2-7B, Method Class=Parameter-Efficient, LoRA Rank=42025.09 | 0.01 | 0.83 | 0 | 0.01 | 0.1 | — | — | 0 | 0.17 | 0 | |
| GA+FILABackbone=Llama2-7B, Training Strategy=Parameter-Efficient, LoRA Rank=rank-162025.09 | 0.01 | 0.83 | 0 | 0.01 | 0.1 | — | — | 0 | 0.17 | 0 |