Knowledge Unlearning on Internal e-commerce benchmark medium-scale seller 387 items (Forget Set)
89.4ROUGEME+GD
Evaluation results
| Method | Links | ||
|---|---|---|---|
| ME+GDBase Model=Llama 3.1 8B2025.05 | 89.4 | 0.1 | |
| BaselineBase Model=Llama 3.1 8B2025.05 | 89.3 | 0.1 | |
| SimNPO+KLBase Model=Llama 3.1 8B, Variant=22025.05 | 45.2 | 0.26 | |
| UnDIAL+KLBase Model=Llama 3.1 8B, Variant=12025.05 | 44.3 | 0.17 | |
| RKLD+KLBase Model=Llama 3.1 8B2025.05 | 29.4 | 0.17 | |
| GA+KLBase Model=Llama 3.1 8B2025.05 | 20.4 | 0.47 | |
| UnDIAL+KLBase Model=Llama 3.1 8B, Variant=22025.05 | 15.1 | 0.27 | |
| NPO+KLBase Model=Llama 3.1 8B, Variant=22025.05 | 13.4 | 0.78 | |
| NPOBase Model=Llama 3.1 8B2025.05 | 13.1 | 1.06 | |
| GABase Model=Llama 3.1 8B2025.05 | 11.6 | 5.03 | |
| NPO+KLBase Model=Llama 3.1 8B, Variant=12025.05 | 11.5 | 0.89 | |
| SimNPO+KLBase Model=Llama 3.1 8B, Variant=12025.05 | 0.3 | 33.63 | |
| Unilogit+KLBase Model=Llama 3.1 8B, Variant=22025.05 | 0.2 | 6.64 | |
| Unilogit+KLBase Model=Llama 3.1 8B, Variant=12025.05 | 0 | 10.78 |