A general model knows a little about many subjects. Fine-tuning teaches it how one kind of work is done. The examples are usually pairs or sequences: a question and the right answer, a customer message and the agent's reply, a draft and its corrected version. Far less data is needed than for pre-training, but it has to be accurate and consistent, because the model will copy the patterns it sees, including the mistakes.
Company data is well suited to this. Resolved support tickets show how a problem was diagnosed and closed. Reviewed documents show a professional standard. Code with review comments shows what a senior engineer accepts. Data for fine-tuning is commonly delivered as JSONL in a chat layout, with each line holding one conversation as a list of messages. In a license, fine-tuning is normally listed as its own permitted use, separate from pre-training and evaluation, and may be priced differently. A seller can grant one use without the others.
Because fine-tuning sets are small, each record carries more weight. Buyers inspect them closely and reject sets with inconsistent formatting, wrong answers or leftover personal details.
Example
A lab fine-tunes a general model on de-identified insurance claim files with their final decisions, so the model learns how adjusters reason through a claim.