Ask HN: Anyone Succesfully fine-tuning LLMs?

Has anyone of you successfully fine-tuned a LLM?

I have done several attempts for simple use cases and the result was always extremely poor generalization.

Any experience, guides, examples would be valuable.