where do you think the corpus of training data comes from?