Investigate using LLMs to generate training data

[marco-c]

E.g. it could be useful for generating datasets of short sentences to help with #215.