The behavior you see is the correct way of the DataRobot Prime model. Compared to the teacher model, the Prime model may be trained only if it has extra data to train the student model. So 100% doesn't leave any data, that the teacher model hasn't seen. So does the 80% teacher model, because holdout data is unreachable during training. So as a result, your error means: "student model doesn't have data to train on"