Confirm that fine-tuning is the right fix
When the problem is missing retrieval context, unclear instructions, or an unstable tool contract, fine-tuning is rarely the first answer. Consider LoRA, adapters, or full tuning only when the task format is stable, examples are sufficient, and behavior truly must change.
Data matters more than parameter count
Deduplicate, redact, and verify rights for the dataset. Split training, validation, and regression samples by time or source so the same question cannot leak. Keep labeling guidance and refusal examples so the model does not learn surface formatting only.
Set a baseline for accuracy, refusal rate, latency, memory, and cost before training. After training, compare gains and regressions in general questions, long context, tool calling, and safety. A single loss value is not an approval to ship.
Release with a rollback path
Save the base model, adapter, training parameters, data version, and evaluation report. Validate with shadow traffic before a small rollout. Limit adapter combinations and watch memory. If safety or quality regresses, switch back to the baseline immediately.
Sources
See Hugging Face PEFT, the Transformers documentation, and the LoRA paper.
