What is Fine-tuning?
Fine-tuning means taking a finished language model and training it further on your own pairs of input and desired output. The result is your own model that already knows your format, your tone and your categories, instead of having to infer them from a long instruction every time. The difference from prompting is permanence: whatever sits in the `system prompt` has to be sent with every call, whereas whatever sits in the model is simply there.
What matters is knowing what fine-tuning is good for. It teaches form, not facts: classification into your categories, an awkward output format, a very particular writing style, dependable behaviour on repetitive tasks. Current knowledge and company data do not belong in the training set. They belong in the model at runtime, through retrieval. A fine-tuned model that memorised March pricing will state June pricing wrongly, and do it convincingly.
The price is effort and lock-in. You need clean training data, a test set that was not part of training, and a plan for the day your provider ships a better base model. Hence the usual order: sharpen the prompt, then put examples in the prompt, then add retrieval, and only fine-tune when that measurably falls short.
Why does Fine-tuning matter?
The bar is lower than most people assume: in its own fine-tuning documentation, OpenAI names 50 to 100 carefully curated examples as a starting point, while explicitly recommending that you exhaust prompting and retrieval first. The bottleneck is not the volume of data, it is whether your problem is a problem of form at all.
Fine-tuning in practice
- 01A support team trains a small model on 300 real ticket classifications so that routing into 12 categories works without a page-long prompt.
- 02An editorial team trains a model on 80 of its own edited articles so first drafts arrive in house style instead of sounding like generic AI.
- 03A sales team needs quote summaries in a fixed JSON structure and trains for exactly that, because the format kept drifting when it lived in the prompt.
