It seems like generating training data with some LLMs that don't have license restrictions on the output should be allowed.