LLM Unlearning on TOFU Llama2-7B (unlearning)
90.8DF ROUGEOriginal
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| OriginalBackbone=Llama2-7B2026.05 | 90.8 | 90.1 | 81.2 | 45.5 | 50.8 | 100 | |
| Fine-tuneBackbone=Llama2-7B2026.05 | 77.2 | 88.7 | 77.5 | 56.3 | 51 | 99.98 | |
| GDiffBackbone=Llama2-7B2026.05 | 41.9 | 60.1 | 76 | 61.4 | 54.1 | 99.98 | |
| DareUBackbone=Llama2-7B2026.05 | 40.7 | 79.7 | 83.4 | 81.6 | 56.4 | 99.62 | |
| RetrainingBackbone=Llama2-7B2026.05 | 38.1 | 91.7 | 81.8 | 100 | 66.9 | 99.71 | |
| SCRNBackbone=Llama2-7B2026.05 | 29.2 | 27 | 65.7 | 25.1 | 77.4 | 99.67 | |
| NPOBackbone=Llama2-7B2026.05 | 28.6 | 31.8 | 47.3 | 23.9 | 59.6 | 99.03 | |
| SCRUBBackbone=Llama2-7B2026.05 | 26.8 | 30.2 | 28.3 | 21.9 | 73.8 | 66.4 | |
| IDKBackbone=Llama2-7B2026.05 | 2.8 | 80.4 | 79.5 | 54.8 | 53 | 100 | |
| GABackbone=Llama2-7B2026.05 | 0 | 0.1 | 0 | 0.9 | 73.9 | 99.9 |