What is fine-tuning? Adapt an existing model with training examples
Understand fine-tuning through support classification, separate parameter updates from prompting and retrieval, and examine data quality, held-out evaluation, and platform availability.
Understand the concept
Fine-tuning continues training an existing model, updating all parameters or selected adaptation parameters for a task. It can teach formats or classification conventions, unlike simply uploading a document to a chat. It does not guarantee recall of every fact. Improve instructions and context first, then compare results on held-out questions to assess useful gains.
Continued training changes model parameters
Examples supplied in a prompt usually provide in-context learning. Fine-tuning uses examples in training and produces updated model parameters or an adapter that later calls load. Both can change responses, but they store and update different things.
For support tickets labeled delivery, refund, or other, begin with clear category definitions and examples. Consider fine-tuning when that remains inconsistent and accurate training examples are available. Learning labels cannot replace querying a real order when its current status is missing.
Training examples should match real inputs
Use the language users actually submit, not only neatly written questions. Resolve cases where annotators assign different categories to the same ticket before training. Training on inconsistent labels does not automatically produce the team’s intended rules.
If production requests contain retrieved passages, reflect a similar context format in training. A single-sentence training question differs substantially from a production request containing pages of evidence. Keep the base-model prompt result to identify which failures training improves.
Check gains and regressions on held-out questions
Correctly answering training questions again only establishes performance on familiar examples. Remove near-duplicates from the test set and include confusing categories. Retain previously successful cases to detect improvements that damage other behavior.
Available models and interfaces depend on the provider. OpenAI’s optimization documentation now warns that its fine-tuning platform is winding down and is unavailable to new users. This page explains the general method; check the chosen platform before implementation. The example was not trained or benchmarked for accuracy or cost.
Checked sources
These sources support the concepts and methods described here. Examples and test scope are labeled separately; check the version of the tools you use.