LLM Unlearning on ArXiv unlearning Llama2-7B
82.5DF ROUGEOriginal
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| OriginalBackbone=Llama2-7B2026.05 | 82.5 | 89.5 | 34.2 | 10.1 | 20.5 | 91.06 | |
| OriginalBackbone=Qwen3-8B2026.05 | 77.9 | 83.5 | 39.2 | 7.1 | 12.6 | 100 | |
| Fine-tuneBackbone=Llama2-7B2026.05 | 64.6 | 76.3 | 48.2 | 6.2 | 17 | 89.83 | |
| NPOBackbone=Llama2-7B2026.05 | 52.2 | 64.2 | 52.8 | 3.9 | 15.4 | 77.74 | |
| Fine-tuneBackbone=Qwen3-8B2026.05 | 49.4 | 58.2 | 53 | 9.3 | 23.1 | 100 | |
| DareUBackbone=Llama2-7B2026.05 | 45.6 | 70.4 | 62.2 | 1.3 | 12.7 | 88.01 | |
| DareUBackbone=Qwen3-8B2026.05 | 40.5 | 65.9 | 66 | 2.5 | 9 | 99.98 | |
| GABackbone=Qwen3-8B2026.05 | 40.4 | 45.8 | 38.5 | 0.6 | 3 | 78.24 | |
| SCRNBackbone=Qwen3-8B2026.05 | 33.7 | 38 | 48.1 | 6.7 | 13.2 | 99.72 | |
| NPOBackbone=Qwen3-8B2026.05 | 30.9 | 38.7 | 50.6 | 3.9 | 15.4 | 99.65 | |
| GDiffBackbone=Llama2-7B2026.05 | 24.3 | 44.2 | 53.2 | 0.1 | 8.1 | 81.83 | |
| GDiffBackbone=Qwen3-8B2026.05 | 20.7 | 43 | 55.8 | 0.1 | 8.1 | 86.36 | |
| GABackbone=Llama2-7B2026.05 | 19.7 | 28.3 | 43.9 | 0.6 | 3 | 65.15 | |
| SCRNBackbone=Llama2-7B2026.05 | 19.6 | 19.1 | 31.2 | 0 | 0.9 | 57.97 | |
| RetrainingBackbone=Llama2-7B2026.05 | 17.2 | 90.8 | 100 | 0 | 19.8 | 57.96 | |
| RetrainingBackbone=Qwen3-8B2026.05 | 17 | 84.6 | 100 | 0 | 14.5 | 68.14 | |
| SCRUBBackbone=Llama2-7B2026.05 | 2.4 | 43.2 | 51.1 | 0 | 6.7 | 2.57 | |
| SCRUBBackbone=Qwen3-8B2026.05 | 2.3 | 34.6 | 46.5 | 0 | 6.7 | 2.54 |