The reasoning steps would also be based on the training set ... it's "predict the training set" all the way down.