Where did most of the code in their training data come from?