Fine-tuning means more training. You show a model many examples of the behavior you want, and the training updates the weights, the numbers the model learned, so that behavior gets more likely. The fast.ai, Hugging Face, and Harvard CS50 AI courses linked above teach how that training step works. A changing handbook usually calls for retrieval, which means putting the current page into the request, not another training run.
What training can change, and what it cannot look up
A fine-tune can push style, format, and a repeated task: “always answer in this JSON shape,” “prefer this tone,” “classify these tickets the way these thousand labels did.” It is a bad place to store a fact that changes. Yesterday’s refund policy, today’s price, and the customer’s order number are data. Data belongs in a retrieval step or a tool. If you bake a price into weights, you have to train again when the price moves, and you still cannot show the page the price came from.
A prompt with three examples is not a fine-tune. People call that few-shot prompting. It spends tokens, the chunks of text the model counts, inside the context window, the text that fits in one request. It leaves the weights alone. Try that before you open a training job. If the examples already get the format right, you may not need a training job.
Training is billed separately from chat
Hosted chat subscriptions and fine-tuning APIs are different products. Inference is a later reply from the trained model. A chat plan does not, by itself, mean you can train the model. When a vendor does sell fine-tuning, the cost is the training run plus later inference, and the data you upload is a policy question. If a salesperson says fine-tuning is included, ask which page says so and save the URL with the date.
A downloadable weight file can still forbid commercial use, or require a separate license for the data. Fine-tuning a weight file you are not allowed to train is a license problem before it is a technical one. The Ollama profile on this site separates the app’s MIT license from the license on each model you pull. Read the model card, the page that comes with the model. The app license does not cover the weights.
A rule that survives a demo
Do not fine-tune to make a model “know our docs.” Put the docs in the request. Do not fine-tune to make a model “take actions.” Give it tools, and decide which actions are allowed. Consider fine-tuning when a prompt and examples still miss a stable pattern, you have enough labeled examples to measure, and someone can say what “better” means on a set the model was not trained on. The courses linked above teach the training step. They do not know your examples.
Sources & limits: This page does not quote a fine-tune price. Check the vendor’s page and save the URL with the date. The Ollama license split is the one on that tool’s profile.