Machine Unlearning on TOFU Forget10 (10% authors split)
0.99Forget Quality - Rouge-LOriginal
Evaluation Results
| Method | Links | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| OriginalBackbone=Llama2-7B2025.09 | 0.99 | 0.51 | 0 | 0.98 | 0.47 | 0.94 | 0.62 | 0.89 | 0.55 | 0.63 | — | — | |
| GDBackbone=Llama2-7B, Tuning Strategy=Full Fine-tuning2025.09 | 0.82 | 0.51 | 0 | 0.92 | 0.47 | 0.92 | 0.6 | 0.88 | 0.55 | 0.62 | — | — | |
| GDBackbone=Llama2-7B, Training Strategy=Full Fine-tuning2025.09 | 0.82 | 0.51 | 0 | 0.92 | 0.47 | 0.92 | 0.6 | 0.88 | 0.55 | 0.62 | — | — | |
| IHLBackbone=Llama2-7B, Tuning Strategy=Full Fine-tuning2025.09 | 0.73 | 0.57 | 0 | 0.88 | 0.45 | 0.94 | 0.64 | 0.89 | 0.59 | 0.64 | — | — | |
| IHLBackbone=Llama2-7B, Training Strategy=Full Fine-tuning2025.09 | 0.73 | 0.57 | 0 | 0.88 | 0.45 | 0.94 | 0.64 | 0.89 | 0.59 | 0.64 | — | — | |
| NPOBackbone=Llama2-7B, Tuning Strategy=Full Fine-tuning2025.09 | 0.53 | 0.65 | 0 | 0.54 | 0.35 | 0.94 | 0.5 | 0.9 | 0.51 | 0.47 | — | — | |
| IHLBackbone=Phi-1.5B, Method Category=Full Fine-tuning Methods2025.09 | 0.53 | 0.49 | 0 | 0.76 | 0.49 | 0.39 | 0.45 | 0.71 | 0.5 | 0.51 | — | — | |
| NPOBackbone=Llama2-7B, Training Strategy=Full Fine-tuning2025.09 | 0.53 | 0.65 | 0 | 0.54 | 0.35 | 0.94 | 0.5 | 0.9 | 0.51 | 0.47 | — | — | |
| GABackbone=Llama2-7B, Tuning Strategy=Full Fine-tuning2025.09 | 0.48 | 0.66 | 0.0024 | 0.49 | 0.33 | 0.93 | 0.51 | 0.91 | 0.5 | 0.39 | — | — | |
| GABackbone=Llama2-7B, Training Strategy=Full Fine-tuning2025.09 | 0.48 | 0.66 | 0.0024 | 0.49 | 0.33 | 0.93 | 0.51 | 0.91 | 0.5 | 0.39 | — | — | |
| KLBackbone=Llama2-7B, Tuning Strategy=Full Fine-tuning2025.09 | 0.46 | 0.65 | 0 | 0.47 | 0.35 | 0.93 | 0.55 | 0.89 | 0.56 | 0.49 | — | — | |
| KLBackbone=Llama2-7B, Training Strategy=Full Fine-tuning2025.09 | 0.46 | 0.65 | 0 | 0.47 | 0.35 | 0.93 | 0.55 | 0.89 | 0.56 | 0.49 | — | — | |
| NPOBackbone=Phi-1.5B, Method Category=Full Fine-tuning Methods2025.09 | 0.45 | 0.61 | 0 | 0.45 | 0.38 | 0.35 | 0.39 | 0.71 | 0.43 | 0.37 | — | — | |
| NPOBackbone=Phi-1.5B, Training Type=Full Fine-tuning2025.09 | 0.45 | 0.65 | 0.0005 | 0.45 | 0.35 | 0.3 | 0.37 | 0.69 | 0.42 | 0.37 | — | — | |
| DPOBackbone=Llama2-7B, Tuning Strategy=Full Fine-tuning2025.09 | 0.44 | 0.55 | 0 | 0.66 | 0.44 | 0.82 | 0.54 | 0.87 | 0.51 | 0.57 | — | — | |
| DPOBackbone=Llama2-7B, Training Strategy=Full Fine-tuning2025.09 | 0.44 | 0.55 | 0 | 0.66 | 0.44 | 0.82 | 0.54 | 0.87 | 0.51 | 0.57 | — | — | |
| DPOBackbone=Phi-1.5B, Method Category=Full Fine-tuning Methods2025.09 | 0.41 | 0.49 | 0 | 0.67 | 0.47 | 0.33 | 0.43 | 0.73 | 0.49 | 0.48 | — | — | |
| IHLBackbone=Phi-1.5B, Training Type=Full Fine-tuning2025.09 | 0.41 | 0.51 | 0 | 0.77 | 0.49 | 0.36 | 0.46 | 0.69 | 0.52 | 0.5 | — | — | |
| Retain90Backbone=Llama2-7B, Retraining Condition=90% of data2025.09 | 0.41 | 0.66 | 1 | 0.98 | 0.47 | 0.92 | 0.61 | 0.88 | 0.55 | 0.63 | — | — | |
| GDBackbone=Phi-1.5B, Method Category=Full Fine-tuning Methods2025.09 | 0.37 | 0.53 | 0 | 0.41 | 0.44 | 0.19 | 0.44 | 0.6 | 0.46 | 0.36 | — | — | |
| GD+SineBackbone=Llama2-7B, Tuning Strategy=Parameter-Efficient, LoRA Rank=rank-322025.09 | 0.32 | 0.53 | 0.87 | 0.92 | 0.47 | 0.94 | 0.62 | 0.89 | 0.6 | 0.68 | — | — | |
| DPOBackbone=Phi-1.5B, Training Type=Full Fine-tuning2025.09 | 0.32 | 0.48 | 0 | 0.76 | 0.48 | 0.32 | 0.43 | 0.72 | 0.49 | 0.48 | — | — | |
| GD+SineBackbone=Llama2-7B, Training Strategy=Parameter-Efficient, LoRA Rank=rank-162025.09 | 0.32 | 0.53 | 0.87 | 0.92 | 0.47 | 0.94 | 0.62 | 0.89 | 0.6 | 0.68 | — | — | |
| LoKUBackbone=Llama2-7B, Tuning Strategy=Parameter-Efficient, LoRA Rank=rank-322025.09 | 0.29 | 0.65 | 0.29 | 0.91 | 0.45 | 0.89 | 0.62 | 0.88 | 0.57 | 0.63 | — | — | |
| LoKUBackbone=Llama2-7B, Training Strategy=Parameter-Efficient, LoRA Rank=rank-162025.09 | 0.29 | 0.65 | 0.29 | 0.91 | 0.45 | 0.89 | 0.62 | 0.88 | 0.57 | 0.63 | — | — | |
| LoKUBackbone=Phi-1.5B, LoRA Rank=4, Method Category=Parameter-Efficient Methods2025.09 | 0.26 | 0.49 | 0 | 0.75 | 0.5 | 0.36 | 0.49 | 0.67 | 0.51 | 0.51 | — | — | |
| GD+SineBackbone=Phi-1.5B, LoRA Rank=16, Training Type=Parameter-Efficient2025.09 | 0.23 | 0.45 | 0.654 | 0.93 | 0.48 | 0.42 | 0.46 | 0.76 | 0.5 | 0.52 | — | — | |
| GD+SineBackbone=Phi-1.5B, LoRA Rank=4, Method Category=Parameter-Efficient Methods2025.09 | 0.22 | 0.48 | 0.942 | 0.9 | 0.48 | 0.42 | 0.45 | 0.75 | 0.49 | 0.52 | — | — | |
| GDBackbone=Phi-1.5B, Training Type=Full Fine-tuning2025.09 | 0.2 | 0.52 | 0 | 0.25 | 0.46 | 0.02 | 0.5 | 0.28 | 0.5 | 0.13 | — | — | |
| LoKUBackbone=Phi-1.5B, LoRA Rank=16, Training Type=Parameter-Efficient2025.09 | 0.13 | 0.56 | 0.0121 | 0.7 | 0.47 | 0.32 | 0.48 | 0.67 | 0.55 | 0.5 | — | — | |
| GD+FILABackbone=Phi-1.5B, LoRA Rank=4, Method Category=Parameter-Efficient Methods2025.09 | 0.12 | 0.65 | 0 | 0.11 | 0.23 | 0 | 0.3 | 0.03 | 0.28 | 0 | — | — | |
| GD+FILABackbone=Phi-1.5B, LoRA Rank=16, Training Type=Parameter-Efficient2025.09 | 0.08 | 0.5 | 0 | 0.09 | 0.22 | 0 | 0.52 | 0.04 | 0.34 | 0 | — | — | |
| GA+FILABackbone=Llama2-7B, Tuning Strategy=Parameter-Efficient, LoRA Rank=rank-322025.09 | 0.02 | 0.86 | 0 | 0.02 | 0.09 | 0 | 0.19 | 0 | 0.18 | 0 | — | — | |
| GA+FILABackbone=Llama2-7B, Training Strategy=Parameter-Efficient, LoRA Rank=rank-162025.09 | 0.02 | 0.86 | 0 | 0.02 | 0.09 | 0 | 0.19 | 0 | 0.18 | 0 | — | — | |
| GD+FILABackbone=Llama2-7B, Tuning Strategy=Parameter-Efficient, LoRA Rank=rank-322025.09 | 0.01 | 0.85 | 0 | 0.01 | 0.09 | 0 | 0.18 | 0 | 0.18 | 0 | — | — | |
| KLBackbone=Phi-1.5B, Method Category=Full Fine-tuning Methods2025.09 | 0.01 | 0.77 | 0 | 0.01 | 0.16 | 0 | 0.24 | 0 | 0.25 | 0 | — | — | |
| GABackbone=Phi-1.5B, Method Category=Full Fine-tuning Methods2025.09 | 0.01 | 0.76 | 0 | 0.01 | 0.15 | 0 | 0.24 | 0 | 0.24 | 0 | — | — | |
| KLBackbone=Phi-1.5B, Training Type=Full Fine-tuning2025.09 | 0.01 | 0.7 | 0 | 0.01 | 0.14 | 0 | 0.41 | 0 | 0.35 | 0 | — | — | |
| GABackbone=Phi-1.5B, Training Type=Full Fine-tuning2025.09 | 0.01 | 0.71 | 0 | 0.01 | 0.14 | 0 | 0.41 | 0 | 0.35 | 0 | — | — | |
| GD+FILABackbone=Llama2-7B, Training Strategy=Parameter-Efficient, LoRA Rank=rank-162025.09 | 0.01 | 0.85 | 0 | 0.01 | 0.09 | 0 | 0.18 | 0 | 0.18 | 0 | — | — | |
| GA+FILABackbone=Phi-1.5B, LoRA Rank=4, Method Category=Parameter-Efficient Methods2025.09 | 0 | 0.35 | 0 | 0 | 0.25 | 0 | 0.38 | 0 | 0.32 | 0 | — | — | |
| GA+FILABackbone=Phi-1.5B, LoRA Rank=16, Training Type=Parameter-Efficient2025.09 | 0 | 0.31 | 0 | 0 | 0.28 | 0 | 0.33 | 0 | 0.43 | 0 | — | — | |
| ASTraining Protocol=Adapters, Training Data Volume Parity=True2025.10 | — | — | — | — | — | — | — | — | — | — | 0.01 | 0.99 | |
| GATraining Protocol=Adapters, Training Data Volume Parity=True2025.10 | — | — | — | — | — | — | — | — | — | — | 0.13 | 0.95 | |
| GA + CETraining Protocol=Adapters, Training Data Volume Parity=True2025.10 | — | — | — | — | — | — | — | — | — | — | 0.2 | 0.96 | |
| GA + KLTraining Protocol=Adapters, Training Data Volume Parity=True2025.10 | — | — | — | — | — | — | — | — | — | — | 0.21 | 0.97 | |
| IHLTraining Protocol=Adapters, Training Data Volume Parity=True2025.10 | — | — | — | — | — | — | — | — | — | — | 0.41 | 0.25 | |
| IHL + CETraining Protocol=Adapters, Training Data Volume Parity=True2025.10 | — | — | — | — | — | — | — | — | — | — | 0.61 | 0.22 | |
| IHL + KLTraining Protocol=Adapters, Training Data Volume Parity=True2025.10 | — | — | — | — | — | — | — | — | — | — | 0.52 | 0.23 | |
| NPOTraining Protocol=Adapters, Training Data Volume Parity=True2025.10 | — | — | — | — | — | — | — | — | — | — | 0.21 | 0.97 | |
| NPO + CETraining Protocol=Adapters, Training Data Volume Parity=True2025.10 | — | — | — | — | — | — | — | — | — | — | 0.04 | 0.93 | |
| NPO + KLTraining Protocol=Adapters, Training Data Volume Parity=True2025.10 | — | — | — | — | — | — | — | — | — | — | 0.32 | 0.78 | |
| ULDTraining Protocol=Adapters, Training Data Volume Parity=True2025.10 | — | — | — | — | — | — | — | — | — | — | 0.34 | 0.41 |