What Is Fine-Tuning? A Plain-English Guide to Teaching an AI Model New Tricks

Close-up of a speaker's brass control knobs and toggle switch on a brushed-metal panel

Photo: Hippopx (public domain). Illustrative image.

A large language model arrives knowing a lot about language in general and nothing about you in particular. It can write a poem, summarise a report, or answer a question — but it will not naturally write in your company’s voice, follow your exact report format, or use your product’s vocabulary. Fine-tuning is the process of teaching it those specifics: extra training, on your own examples, on top of the training it already had.

Think of it like the knobs on a stereo. The factory settings play everything reasonably well. Fine-tuning is adjusting the dials for your room, your speakers, your music — small changes, carefully made, that make the same machine sound right for you. Turn the wrong knob too far and the sound gets worse, not better. The same is true of models.

Frequently asked questions

What exactly is fine-tuning?

Fine-tuning takes a pre-trained model and continues training it on a smaller, focused dataset of examples of the behaviour you want. Each example is typically a pair: an input and the output you want for that input. A customer-support company might use hundreds of real support conversations, each with the reply its best agents gave. A law firm might use contracts marked up the way its partners like them. The model adjusts its internal weights slightly, in the direction of those examples, so that its future outputs look more like them.

It is worth being precise about what changes. Fine-tuning adjusts how the model responds — its style, format, tone and habits. It does not reliably teach it new facts. If you fine-tune a model on last year’s product manual, it will not suddenly know this year’s prices; it will just answer in the manual’s style. New knowledge belongs in retrieval, not in fine-tuning.

How is it different from prompting?

Prompting is telling the model what to do in the moment: instructions and examples included in each request. Fine-tuning bakes the behaviour into the model itself, so you do not have to repeat the instructions every time. The trade-off is flexibility against reliability. A prompt can be changed in seconds; a fine-tuned model takes time and data to change, but it behaves the same way every time, uses fewer tokens per request, and cannot have its instructions accidentally dropped or overwritten by a long conversation.

What is LoRA, and why does everyone mention it?

Full fine-tuning updates every weight in the model — expensive, slow, and it needs serious computing power. LoRA (Low-Rank Adaptation) is a cheaper method that freezes most of the model and trains only a small set of extra weights bolted on the side. The result is a small file — often a few hundred megabytes instead of tens of gigabytes — that can be swapped in and out. It is the method most people actually use, and it is why open models ship with fine-tuning code: Runway’s Praxis-1 robot model, released this week, ships with both full and LoRA fine-tuning code, so anyone can adapt it to their own robot.

When should you fine-tune, and when should you not?

Fine-tune when you have a repeated task with a consistent correct style, enough good examples to teach it — usually hundreds at minimum — and evidence that prompting alone is not reliable enough. Good candidates: support replies in a fixed tone, classification into your categories, code in your codebase’s conventions, extraction into a strict format.

Do not fine-tune when the task changes often, when you have only a handful of examples, when the model needs fresh facts, or when a good prompt already works. And do not fine-tune to fix a model that is wrong about the world: a model that hallucinates will hallucinate in your style instead. If the problem is knowledge, use retrieval — giving the model the right documents at answer time — rather than training.

What does it cost?

More than prompting, in two ways. There is the training cost: computing time, which the provider charges for, and which grows with the size of the model and the number of examples. And there is the running cost: most providers charge more per token for a fine-tuned model than for the base model. There is also a hidden cost — maintenance. Every time the base model is updated, or your needs change, the fine-tuned version may need retraining. Budget for that before you start, not after.

What can go wrong?

Three things, most commonly. Overfitting: the model memorises your examples instead of learning the pattern, so it performs well on your training data and badly on anything new. The fix is more varied examples and holding some back to test on. Catastrophic forgetting: heavy fine-tuning can damage what the model already knew, so it gets worse at general tasks. The fix is gentler training — lower learning rates, fewer steps, methods like LoRA that touch less of the model. Baked-in bias: the model learns your examples faithfully, including their mistakes, gaps and prejudices. If your support replies were curt with one group of customers, the fine-tuned model will be too. The quality of the output is capped by the quality of the examples.

How do you know it worked?

Test it the way you would test anything else: on examples it has never seen, scored against the behaviour you actually want, and compared with the plain prompted model on the same test. If the fine-tuned version is not clearly better on unseen examples, it has not earned its cost. Enterprise tools are now building this in — ElevenLabs’ new agent upgrades include fine-tuning models on live customer data — but the discipline is the same whatever the tool: measure before you trust.

The bottom line

Fine-tuning is a dial, not a magic wand. It is the right tool when you know exactly how a model should behave, you have the examples to teach it, and prompting has proved too unreliable or too expensive at scale. For most people, most of the time, the order is: prompt first, add retrieval when the model needs facts, and fine-tune last — when you have proved the task matters and you have the data to teach it. Judge any fine-tuning claim the way you judge any AI claim: what does it do repeatedly, on examples it has never seen, and what does it quietly get wrong?

Sources

  • fine-tuning
  • AI models
  • machine learning
  • explainer
  • how-to

Related articles

Guides

How to Evaluate a New AI Tool Before You Trust It

New AI tools launch every week, and every launch claims to change everything. A practical checklist for testing a tool on your own real tasks before you pay, switch, or rely on it.

· AI tools, buying guide, productivity, how-to

Guides

AI Hallucinations: Why Chatbots Make Things Up — and How to Catch Them

Chatbots invent books, misquote laws and state wrong dates in a perfect, confident tone. That behaviour is called hallucination, and it follows directly from how language models work. Why it happens, the forms it takes, and a working routine for catching it.

· hallucinations, chatbots, explainer, fact-checking, AI basics

Guides

How to Spot AI-Generated Photos, Videos and Voices: A Practical Guide

AI fakes are good enough that 'it looked real' is no longer a test. A practical checklist for photos, videos and voice clips: the details that still give generators away, the provenance tools that check for you, and the one habit that beats them all.

· deepfakes, AI detection, media literacy, how-to, verification