Fine-tuning
Continuing training on task examples so a model reliably produces the behavior, format, or tool use an institutional job requires.
Fine-tuning continues a model's training on examples of the work it has to do. Supervised fine-tuning — the common form — shows the model input and the desired output, repeatedly, until the mapping holds.
What it earns its keep on
Format consistency is the clearest case. When the output has to match a fixed schema, a regulated document layout, or a report template that gets rejected on deviation, fine-tuning holds that format across volume in a way prompting does not. Tool-use contracts and domain vocabulary are the next clearest.
What it is the wrong tool for
Raw knowledge. Fine-tuning a model on a corpus of operational documents bakes those documents into weights that cannot be updated without another training run, and cannot tell an operator which answer is now stale. Even where it edges out retrieval on a held-out set, the first document change makes the model wrong with no signal. Knowledge that changes belongs in retrieval.
What it costs
A fine-tuned model is a new artifact. It needs its own evaluation, its own version, its own approval, its own monitoring, and its own rollback path. Training data rights, base-model license terms, and privacy boundaries all have to permit the training before it starts. Those obligations are the real cost, not the compute.
The honest default
Try the smallest thing first: a careful prompt, a structured output contract, and a validator. Reach for fine-tuning when a representative evaluation shows that baseline missing a gate the work actually has — not because the model could be better in general.
Often confused with
How it differs from similar terms.
| fine-tuning | Compared with | What's different |
|---|---|---|
| Supervised fine-tuning | Preference optimization | SFT teaches the mapping from input to a correct output. Preference optimization teaches which of two outputs is better. |
| Fine-tuning | Prompting | Prompting is free and reversible. Fine-tuning is a versioned artifact with evaluation, approval, and rollback obligations. |
Want these terms worked through on one of your own decisions? Start with a Decision Blueprint