What is fine-tuning? Explained with specialist training after school

In one sentence: Fine-tuning is a short, extra round of training that takes an AI model that already knows a lot in general and teaches it one specific job, using a set of example questions paired with the answers you want.

The everyday picture

Think of a bright graduate who joins your office. School gave them a broad education. They can read, write, reason and talk about almost anything. But they have never written a letter in your firm's style or sorted your kind of paperwork.

So you send them on a short specialist course. They do not go back to school. Instead, they study a folder of worked examples from your office: this request, this ideal reply. After the course, they do that one job faster and more consistently, and they still have everything school taught them.

An AI model works the same way. The big general training is the school. Fine-tuning is the specialist course that comes after it.

How it works, step by step

1. The base model (the graduate) has already been trained on a huge amount of general data by an AI company. It is good at many things but knows nothing about your way of working.
2. The training examples (the folder of worked examples) are pairs of a typical request and the ideal response, written or checked by people who know the job. OpenAI's guide asks for at least 10 and suggests starting with about 50 well-made examples.
3. Fine-tuning (the specialist course) runs those examples through the model and nudges its internal settings, called weights, so its answers move closer to yours. A popular method called LoRA leaves the original weights frozen and trains a small add-on instead, which makes the course much cheaper.
4. The tuned model (the newly qualified specialist) is saved as a separate model with its own name. The original base model is not changed.
5. Testing (the course's final test) compares the tuned model with the original on examples it never saw during the course. Only if it does better does it go to work in your app.

Where the picture stops

• A graduate understands why your examples are right. A model only picks up patterns from them. With too few or too similar examples it can simply memorise them, which OpenAI's guide warns about, and then do worse on new requests.
• A specialist can read this week's price list and use it. A tuned model is fixed when the course ends. If facts change, you would have to run the course again, so changing facts are usually better handed to the model with each question instead.
• A person keeps their training for life. A tuned model is built on one specific base model, and when the provider retires that base model, the tuned version is retired with it.

Why it matters for your business

• Consistency: OpenAI lists sorting requests into categories, producing a set format, nuanced translation and fixing cases where the model ignores instructions as good uses. These are jobs where the same shape of answer is needed every time.
• Effort and cost: you need good examples, a way to test results, and someone to run the job. OpenAI's guide says to set up tests before investing in fine-tuning at all, and to rethink the task or the instructions if about 50 examples make no difference.
• Risk of lock-in: a tuned model depends on its provider and its base model. OpenAI, for example, says on its guide that its fine-tuning service is no longer open to new users and is being wound down. Check what your provider offers today, and how long it will last, before you plan around it.
• Cheaper ways to tune: methods like LoRA cut the number of trained settings by about 10,000 times compared with fully retraining a very large model, and to need about three times less computer memory, according to its authors. That makes tuning far more affordable.

An example

Imagine a 30-person insurance broker that receives hundreds of emails a day. Staff want each one tagged as a claim, a quote request, a complaint or a policy change. A general AI model gets most of them right, but it keeps inventing extra categories. The broker's team first tries clearer instructions and tests the results. It is still not reliable enough. So they collect a few hundred past emails, each tagged correctly by an experienced colleague, keep some aside for testing, and fine-tune a model on the rest. The tuned model uses only the four tags, and the held-back emails show whether it really sorts better than before.

Frequently asked questions

Is fine-tuning the same as training a model from scratch?

No. Training from scratch is the long general schooling. Fine-tuning is a short extra course on top of a model that is already trained, using far fewer examples.

Can fine-tuning teach an AI model my company's documents?

It is better at teaching a style, a format or a task than at storing facts that change. For answering from your documents, giving the model the relevant pages with each question, called RAG, is usually easier to keep up to date.

How many examples do I need to fine-tune a model?

It varies by task. OpenAI's guide needs at least 10 and suggests starting with about 50 good examples, then adding more only if tests show it helps.

Does fine-tuning change the original model?

No. It creates a new, separate model that you call by its own name. The base model stays as it was.

Related explainers

• Training vs inference
• What is RAG?
• What is a system prompt?
• What is an LLM?

The DevDrift builds AI automation and helps teams test whether better prompts, RAG or fine-tuning fits the job: see how.

Sources

• OpenAI: Supervised fine-tuning guide
• Hu et al.: LoRA: Low-Rank Adaptation of Large Language Models (arXiv)