Is it possible to fine-tune gemma4 A4B to generate complex legal principles of court decision? [D]

I have a big database of local court decisions with a legal sentence which is like a paragraph summary of the doc. I've been tinkering with FTing for many days now, all results inconclusive never beating base except for a highly specific task where the eval was built around a specific task of extracting passages from the text, where it seemed to match the gold, but as I've found out it is unusable and a lot of padding and generalizations which I need to actually eliminate. So the question is, can a 26B model even be fine-tuned to produce those complex legal principle/headnote of a court decision? If so how would you do it? I built a LLM eval, then used a standard unsloth UI on a rented server and fed it the most HQ training data from the whole set (100k decisionswith a legal sentence), I fine-tuned both base and IT variant of Gemma 4 26B A4B. Neither beat a prompted "base" model on my evals. I also used Claude fable 5 to vibe code the whole project, could it be that I have made some trivial errors because of it? I know I'm not giving you much context, but as an expert in the field does this sound difficult or doable? Any gotchas that stand out to you immediately? Thanks for reading