I’m totally ignorant about LLMs for specific applications, but why don’t you actually try and get a feel for which language model to use as a base?
In general at HF, it would be quickest if you could find a space that uses LLM for a similar use case and see the source code there or ask the author a question.