gemma-finetune

QLoRA fine-tune a Gemma 4 family model on an instruction dataset, save the LoRA adapter, run a baseline-vs-tuned eval, and emit a compare.md showing observable behavior shift. Recipe layer — assumes you already have a CUDA GPU (use nebius-gpu skill for infra).

Provision a GPU VM on Nebius AI Cloud, run a workload via SSH, then tear down cleanly. Use when user asks to "spin up a GPU on Nebius", "fine-tune on Nebius", "provision an H100/H200/L40S", or sets up Nebius for the first time. Cost-safe — every flow ends in mandatory teardown.