Extractive Question Answering on MLQA Arabic (test)
86.18BERTScore F1GRASP LoRA Merge + pruning
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| GRASP LoRA Merge + pruningAdaptation Method=GRASP LoRA Merge, Training Languages=Arabic+English, Pruning Strategy=Learned prune ratio2026.01 | 86.18 | 48.97 | 41 | 52.51 | |
| Adaptive LoRA Merge + pruningAdaptation Method=Adaptive LoRA Merge, Training Languages=Arabic+English, Pruning Strategy=Best grid search over [10% – 80%]2026.01 | 85.62 | 40 | 38.33 | 50.29 | |
| Adaptive LoRA MergeAdaptation Method=Adaptive LoRA Merge, Training Languages=Arabic+English, Pruning Strategy=Without Pruning2026.01 | 84.8 | — | 36 | 47.77 | |
| LoRAAdaptation Method=LoRA, Training Languages=Arabic, Pruning Strategy=Without Pruning2026.01 | 84.07 | — | 29 | 43.84 | |
| LoRAHubAdaptation Method=LoRAHub, Training Languages=Arabic+English, Pruning Strategy=Without Pruning2026.01 | 83.46 | — | 29 | 42.79 | |
| TIES mergeAdaptation Method=TIES merge, Training Languages=Arabic+English, Pruning Strategy=Without Pruning2026.01 | 67.19 | — | 2 | 13.28 | |
| Zero shotAdaptation Method=Zero-shot, Training Languages=Arabic, Pruning Strategy=Without Pruning2026.01 | 65.71 | — | 0 | 11 |