Fine-tune bake-off: LFM2.5-350M vs Gemma-3-270M/FunctionGemma vs Qwen2.5-0.5B-Instruct

[espetro]

LoRA/QLoRA (peft/trl or unsloth) on identical data, cloud/rented GPU. Same recipe as ScribeAgent / distil labs LFM2.5 fine-tune.