There's a lot of reasoning in the training data.