Extractive Question Answering on MLQA Chinese (test)
71.28BERTScore F1GRASP LoRA Merge + pruning
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| GRASP LoRA Merge + pruningAdaptation Method=GRASP LoRA Merge, Training Languages=Chinese+English, Pruning Strategy=Learned prune ratio2026.01 | 71.28 | 23.73 | 48 | 51.6 | |
| Adaptive LoRA Merge + pruningAdaptation Method=Adaptive LoRA Merge, Training Languages=Chinese+English, Pruning Strategy=Best grid search over [10% – 80%]2026.01 | 69.3 | 10 | 46.5 | 50.93 | |
| Adaptive LoRA MergeAdaptation Method=Adaptive LoRA Merge, Training Languages=Chinese+English, Pruning Strategy=Without Pruning2026.01 | 67.45 | — | 46 | 49.63 | |
| LoRAHubAdaptation Method=LoRAHub, Training Languages=Chinese+English, Pruning Strategy=Without Pruning2026.01 | 65.11 | — | 44 | 48.1 | |
| LoRAAdaptation Method=LoRA, Training Languages=Chinese, Pruning Strategy=Without Pruning2026.01 | 64.15 | — | 42 | 47.17 | |
| TIES mergeAdaptation Method=TIES merge, Training Languages=Chinese+English, Pruning Strategy=Without Pruning2026.01 | 11.28 | — | 0 | 13.51 | |
| Zero shotAdaptation Method=Zero-shot, Training Languages=Chinese, Pruning Strategy=Without Pruning2026.01 | 7.12 | — | 0 | 6.95 |