Supervised fine-tuning (SFT)
Training a model on pairs of an input and the expert-approved output, so it does one task in the institution's own format.
Category
How a model is chosen and adapted
Related terms
Supervised fine-tuning (SFT) trains a model on pairs of an input and the output an expert approved for it, so the model learns to do one task in the institution's own format and vocabulary.
Where the examples come from
Past cases the institution has already decided: an accepted memo section, a field extracted correctly, a reply a team lead approved. Personal data is removed first. The examples are kept apart from the test set the customer's expert grades before training, so the model is measured on cases it never saw.
What it changes and what it does not
It changes behavior: format, structure, wording, and when to say that information is missing. It does not keep changing knowledge current, which is retrieval's job. It grants no authority either: what may be approved stays in the institution's written rules. The model goes live only on the tasks it passes.
The outsourced complaints desk use case shows it: each client's wording is taught by fine-tuning instead of a prompt.
Often confused with
How it differs from similar terms.
| supervised fine-tuning (SFT) | Compared with | What's different |
|---|---|---|
| Supervised fine-tuning | Fine-tuning | Fine-tuning is the general family. Supervised fine-tuning is its common form: it learns from approved input and output pairs. |
Want these terms worked through on one of your own decisions? Start with a Decision Blueprint