LLM Unlearning on TOFU (Metrics: DF ROUGE, DR ROUGE, ToW, Truth Ratio, MIA)
93.2DF ROUGEOriginal
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| OriginalFull-parameter unlearning=true, LoRA=false, Backbone=Llama2-7B, Seed=422026.05 | 93.2 | 92.6 | 92.9 | 37.9 | 40.1 | 99.6 | |
| OriginalBackbone=Qwen3-8B2026.05 | 84.8 | 80.8 | 66.4 | 50.1 | 37.8 | 100 | |
| Fine-tuneBackbone=Qwen3-8B2026.05 | 72.7 | 82.7 | 71 | 70.2 | 39 | 100 | |
| NPOBackbone=Qwen3-8B2026.05 | 64 | 63.7 | 70.4 | 63.8 | 47.2 | 98.92 | |
| SCRNBackbone=Qwen3-8B2026.05 | 57.3 | 55.4 | 45.7 | 40.9 | 26.2 | 94.53 | |
| SCRUBBackbone=Qwen3-8B2026.05 | 49.8 | 64.4 | 68.9 | 73.1 | 42.5 | 99.42 | |
| DareUBackbone=Qwen3-8B2026.05 | 46.5 | 71.8 | 67 | 79.7 | 40.6 | 99.99 | |
| RetrainingBackbone=Qwen3-8B2026.05 | 37.5 | 82.7 | 68.7 | 100 | 53.7 | 99.67 | |
| RetrainingFull-parameter unlearning=true, LoRA=false, Backbone=Llama2-7B, Seed=422026.05 | 34.1 | 99.4 | 97.2 | 100 | 99.5 | 96.2 | |
| DareUFull-parameter unlearning=true, LoRA=false, Backbone=Llama2-7B, Seed=422026.05 | 28.8 | 79.3 | 77.8 | 75.2 | 76.8 | 97.8 | |
| GDiffBackbone=Qwen3-8B2026.05 | 28.1 | 34.7 | 52.7 | 40.6 | 41.4 | 70.29 | |
| IDKBackbone=Qwen3-8B2026.05 | 19.1 | 75.4 | 70.5 | 66.8 | 40.6 | 100 | |
| GABackbone=Qwen3-8B2026.05 | 0.3 | 0.3 | 0 | 2.8 | 51.2 | 38.29 |