Pre-trained model:
Stacked adapter on top of Language adapter. MAD-X 2.0 style. The language adapters in the last layer (layer 11) are deleted.