I’m starting to think that the problem with my fine tuning experiments was that I had not found a LLM dumb enough to actually be able to improve it 😅