What happens to a model during fine-tuning?
A model is a very large set of numbers, fixed during its original training. Fine-tuning carries that training on for a short while with a small, focused set of examples: pairs of "here is the input" and "here is exactly the output I want". Each pass nudges the numbers so the model's own output drifts towards your examples. What comes out is a new copy of the model that writes in the trained way by default.
A 2021 method from Microsoft researchers called LoRA (low-rank adaptation) made this far cheaper, because only a small add-on is trained while the original model stays as it was.
Prompt, look-up or fine-tune: which does the job?
| What you want | Try first | Why |
|---|---|---|
| Your prices and policies used correctly | Put them in the prompt, or in a document the AI looks up | Facts change, and a trained-in price cannot be edited |
| Writing in your voice | A prompt with 3 to 5 of your best examples | Usually close enough, and free to change |
| The same tightly formatted output, thousands of times | Fine-tuning | More consistent format, shorter prompts on every call |
| Last week's sales figures | Give it the figures | No training can know what happened after it |
How many examples does fine-tuning need?
More than most owners have ready. OpenAI's supervised fine-tuning guide sets a minimum of 10 examples, says improvements usually show at 50 to 100, and suggests starting with 50 well-made ones; if 50 make no difference, it says to rethink the task or the prompt before adding more. Every example has to be right, because the model learns your mistakes as faithfully as your good work.
The same guide now says OpenAI is winding down its fine-tuning service and no longer takes new users. Other providers still offer it, and open-weight models can be fine-tuned by anyone with a capable computer.
What would teaching a model its voice cost a saddlery? (example figures)
A saddlery in Rockhampton sells 600 products online and wants every description in its own plain, practical voice, the way the owner talks to a stockman across the counter. There are two routes.
| Fine-tuning | A prompt with examples | |
|---|---|---|
| Owner's time | 100 example pairs at 15 minutes each: 25 hours | A one-page voice guide and 5 best descriptions: 2 hours |
| At $60 an hour | $1,500 | $120 |
| Outside help | A developer to prepare, train and connect it: about $2,000 | None |
| When the model is retired | Train again | Paste the same guide into the new one |
| Cost to start | About $3,500 | $120 |
The sensible order is the cheap one first. Write the voice guide, run it on 20 products and read the results. If 18 of the 20 need only light edits, the saddlery is done. Fine-tuning earns its cost only when the prompt route keeps failing the same way across hundreds of items.
Where does CoraVision get the specifics of your business?
CoraVision gathers the specifics of your business by asking and by reading your accounts. The onboarding call (about 25 minutes, nothing to prepare) covers what you are aiming for and what a typical sale is worth, and Cora works out what a customer is worth over time from those answers. Once Google Ads, Meta and Google Analytics 4 are connected, CoraVision reads them every night. From all of that, Cora builds a 12-month strategy around one number you choose to chase.
Questions
Can fine-tuning teach an AI my price list?
Badly. Facts trained into a model cannot be edited when they change, and the model may blend old and new figures. Keep prices in a document the AI looks up, or paste them into the prompt, so that changing a price means changing one line.
Is fine-tuning the same as training a model from scratch?
No. Training a large model from scratch takes months of computing in a data centre. Fine-tuning starts from a finished model and adjusts it with a small set of examples, often in a matter of hours.
Is a custom GPT fine-tuned?
No. OpenAI's custom GPTs are standard models given saved instructions, uploaded files and optional connections to other tools. That is a prompt and a look-up packaged together, which is why anyone can build one in an afternoon.