Flag if see [A, B, C] in training then [A, B] or [A, B, C, D] in evaluation. Would go in `target.py` I think, and have sklearn component.