“A category of synthetic data generation involves producing text with language models that can be easily and automatically verified, such as in mathematics or for LeetCode-style coding problems with test cases.”