Knowledge Retention on internal e-commerce benchmark large-scale seller 1065 items (Neighbours)
58.2Rouge ScoreME+GD
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| ME+GDBackbone=Llama 3.1 8B2025.05 | 58.2 | 0.54 | |
| BaselineBackbone=Llama 3.1 8B2025.05 | 58.1 | 0.54 | |
| Undial+KLBackbone=Llama 3.1 8B, variant=12025.05 | 55.8 | 0.59 | |
| Unilogit+KLBackbone=Llama 3.1 8B, variant=22025.05 | 55.6 | 0.6 | |
| GA+KLBackbone=Llama 3.1 8B2025.05 | 52.5 | 0.62 | |
| NPO+KLBackbone=Llama 3.1 8B2025.05 | 51.7 | 0.7 | |
| Undial+KLBackbone=Llama 3.1 8B, variant=22025.05 | 50.2 | 0.64 | |
| Unilogit+KLBackbone=Llama 3.1 8B, variant=12025.05 | 46.8 | 0.69 |