Linguistic Probing on MultiBLiMP (test)
93.3Accuracy (CA)Full FT
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| Full FTAdaptation Strategy=Full Finetuning2026.05 | 93.3 | 81.9 | 81.7 | 92.7 | 91.9 | 96.5 | |
| AEFTAdaptation Strategy=All Expert Finetuning2026.05 | 89 | 74 | 74.6 | 89.9 | 85.5 | 94.4 | |
| SSFTAdaptation Strategy=Selective and Shared Expert Finetuning2026.05 | 85.4 | 75.3 | 73.7 | 88.6 | 85.1 | 93.3 | |
| SEFTAdaptation Strategy=Selective Expert Finetuning2026.05 | 82.9 | 68.4 | 66.5 | 86.3 | 81.7 | 86.5 | |
| M7Adaptation Strategy=M7 Baseline2026.05 | 81.3 | 61.5 | 68.5 | 86 | 80.7 | 83.3 |