Training data, from what a model gets wrong.

Four models write questions, answer them and score your model’s attempt. The ones it gets wrong, with reference answers and a rubric, become its training set.