Fine-tuning

Fine-tuning takes a trained model and runs more training on examples you supply, usually pairs of inputs and ideal outputs. Unlike prompting, it changes the model itself. The result is a model, or a small adapter on top of one, that does your task well with far less instruction in the prompt.

For product features, fine-tuning is rarely the first move. Better prompts, few-shot examples, retrieval, structured outputs, or a stronger model fix most quality problems faster and keep you portable. Fine-tuning earns its place when the task is narrow and stable, you have plenty of good examples, outputs can be scored automatically, and something pushes you toward a smaller model: per-token cost at volume, latency, or data that must stay on your hardware.

Two misconceptions waste the most effort. Fine-tuning is a poor way to teach a model facts that change; that's what retrieval is for. And it can't conjure a capability no model has; it mostly sharpens behavior a model can already approximate. Hosted fine-tuning offerings also come and go, while an open-weight model you tune yourself doesn't depend on any provider's roadmap. Whichever route you take, hold out a test set before training and compare against your best prompt-only baseline.