SciLad encoder models (like BERT/Roberta) for task-specific fine tuning.