Maybe using a larger model to generate synthetic data of question path Combos, and also to rephrase and generate similar type questions for a more varier training set.