What is fine-tuning?
Teams pay to retrain a model hoping it will learn last week's prices, then find it inventing them just as confidently as before.
Fine-tuning
supervised fine-tuningSFT
Fine-tuning trains an existing base model further on your own paired examples of inputs and outputs, producing a version that does your task by default. It shapes how a model responds. It does not give the model new facts to look up.
OpenAI describes the shape of it plainly. Take a base model, give it the inputs and outputs you expect in your app, and get a model that does well at those tasks. You supply the pairs. The training run does the rest.
The result is a private model that behaves your way with a much shorter prompt. Where you once pasted six examples and a page of rules on every request, the behaviour is now baked in.
The cases where it earns its cost
Try prompting first. That is OpenAI's own guidance, which says the prompt engineering process may be all you need. Unusual advice from a vendor selling the training run, and it is right. The case for training arrives further along.
OpenAI names four situations where it pays. You have more good examples than fit in one request. You want shorter prompts, to cut tokens and latency at scale. You want to train on private data without sending it on every call. Or you want a smaller model to handle one task well.
Read that list again. Every item is about form, cost or scale. None is about knowledge.
- CollectReal input and output pairs.
- CleanBad examples teach bad habits.
- Hold backA test set it never sees.
- TrainA private model is produced.
- EvaluateCompare against the base model.
Skip the held-back set and you cannot tell improvement from memorisation. That is the step that gets cut when a deadline moves.
Related questions
01How many examples does fine-tuning need?
More than most teams expect, and quality decides the outcome more than quantity. A few dozen clean, consistent examples of one narrow task beat a thousand scraped from mixed sources. Contradictory examples teach the model to be inconsistent.
02Can fine-tuning make a model stop hallucinating?
No, and it can make it worse. Training on examples of confident answers teaches confident phrasing, not accuracy. Grounding answers in retrieved sources is what reduces invention, because the facts arrive with the question rather than from memory.
03What do you lose by fine-tuning?
Flexibility and portability. You now own a model version that must be re-run when the base model retires. Prompt changes no longer fully control behaviour. Budget for that upkeep before the first training run.
