at present predict_proba assumes that the second of the classes is the target class. so for 0, 1 it assumes 1 for 1,2, it will use 2. this should be generalized into a configurable option to be given to gen_workflow.