Quick answer: Fine-tuning takes an AI model already trained on a huge amount of general material and gives it a short extra round of training on a small, focused set of examples, so it gets better at one specific job.
The same generative AI model that can draft a sonnet and explain photosynthesis will often fumble one simple question about your company’s return policy. It knows a great deal in general and nothing about you in particular. Fine-tuning is the step that closes that gap.
This page covers how that works, what it is used for, how it differs from the two terms it keeps getting confused with, and whether you need it.
What Is Fine-Tuning in AI?
Fine-tuning is what happens when an AI model that already knows a lot in general takes a short extra course in one specific subject.
Think of a new hire with a good education. They read, write, and reason well, but they have never seen your products. You do not send them back to university. You hand them a few hundred examples of good answers.
Pretraining is the general education. The fine-tuning dataset is the stack of examples. The person who walks out is the same person who walked in, just better at this one job.
More precisely, fine-tuning continues training a pretrained model on a smaller, targeted dataset so it fits one specialized use case while keeping most of its general ability. Pretrained means it was trained once, expensively, on a very large general collection of text or images. That is how almost every large language model was built. Fine-tuning is a form of transfer learning, using what a model learned on one task to get better at a related one.
Fine-tuning is not building a new model from scratch, since the expensive part is already done. It is also not the same as writing better instructions for a model you already have. That has its own name, and a later section sorts all three apart.
How Does Fine-Tuning Actually Work?
Fine-tuning data is less exotic than it sounds. It is mostly questions paired with the answers you wish the model had given.
A dataset is usually a plain text file where each line holds one example: an input and the output you want back. For example, a small swimwear shop might pair “Can I return a swimsuit?” with its own policy: unworn, hygiene liner intact, within 30 days. A few hundred lines like that teach the shop’s rules and tone.
- One line, one example. Each pairs an input with the response you want.
- A few hundred clean examples is a normal start. Roughly 200 to 1,000 for a narrow task, not millions.
- Consistency beats volume. Three hundred careful examples outperform three thousand sloppy ones.
Inside the model, nothing dramatic happens. A model is a very large set of internal settings, tuned once during pretraining, that decide how each word follows the last. Fine-tuning nudges some of those settings gently with your examples, which is why it is fast and cheap next to the original training run.
The early parts of a transformer model, which handle basic language sense, are left alone. The later parts, which shape the answer, get adjusted.
You can retune the whole model, or bolt on a small extra piece and train only that. The second is the common modern route, and you will meet its name: LoRA, or low-rank adaptation. It is cheaper, it runs on rented cloud hardware, and it disturbs the model’s original abilities less. The original LoRA research reported cutting the number of parameters that need training by a factor of about 10,000 while matching the quality of retuning everything.
Fine-tuning is a small second pass, not a rebuild.
What Is Fine-Tuning Used For?
Fine-tuning is mostly used for narrow, unglamorous jobs, not for making a model smarter overall. A fine-tuned model is usually worse than the original at everything except the one thing it was tuned for. That trade is the whole point.
Three patterns cover most of it.
Matching a Company’s Voice and Rules
A business feeds the model its own past support conversations, product details, and policies, so answers come back reflecting its actual rules instead of generic advice. The tone half matters as much as the knowledge half. Instructions can tell a model to sound warm and brief. A few hundred real examples show it what that sounds like in your business.
Sorting and Labeling Text at Scale
Classification is the textbook fit: sentiment analysis, routing incoming messages into categories, flagging which support tickets are urgent. The task is narrow, repetitive, and has one right answer, so a few hundred labeled examples go a long way. These are also the cheapest fine-tuning jobs to run, because the outputs are short and the datasets stay small.
Specialized and Small-Scale Projects
Not all of this happens at enterprise scale. An independent developer or a two-person team might tune a small open model to write in one house style, read a specialized document format, or handle the vocabulary of a single trade that general models keep getting wrong. A project like that is within reach of one person with rented cloud time.
The common thread is a narrow job with one right answer. That is also the test for whether fine-tuning is your answer at all.
Fine-Tuning vs. Prompt Engineering vs. RAG: Which One Do You Actually Need?
These three get used as if they were rival products. They solve different problems, and picking wrong costs weeks.
Prompt engineering changes only the instructions you give when you ask, so it is instant, free, and limited to what the model already does. RAG, or retrieval-augmented generation, connects the model to an outside store of documents so it can look facts up while answering. Fine-tuning changes the model itself, permanently, and suits behavior rather than facts.
| Approach | What it changes | Best for | Effort and cost |
|---|---|---|---|
| Prompt engineering | The instructions you send | Behavior it can already produce | Minutes, near free |
| RAG | What the model can see | Facts that are yours or change often | Days, some setup |
| Fine-tuning | The model itself | A narrow, stable, repeated task | Weeks, mostly your time |
Try better instructions first. It is fastest and solves more than people expect. If the model fails because it does not know your facts, or those facts keep changing, that is retrieval, not training. People also reach for fine-tuning reflexively, on problems it does not solve: off-format output, confidently wrong facts.
Only reach for it once prompting and retrieval have plateaued, you can write a few hundred clean examples of the behavior you want, and you can measure whether the result is better. The real cost is rarely compute. It is maintaining the example set.
For most people asking what fine-tuning is, the answer is that they will never need to do it.
What Can Go Wrong When a Model Is Fine-Tuned?
Teaching a model new facts can make it worse at facts it already had right. Researchers studying this measured a drop on the order of 15% on held-out questions unrelated to the new material, an effect they call factual forgetting.
- Forgetting. Beyond facts, the model can lose or blur general abilities it had before, a problem practitioners call catastrophic forgetting. Small add-on approaches like LoRA reduce it.
- Overfitting. Train on too few or too narrow a set of examples and the model learns the examples themselves rather than the pattern behind them, so anything slightly different throws it.
- New hallucinations. Tuning that pushes new facts in can leave a model confidently wrong about things it previously got right, and the effect is worse when the new information resembles what it already knows. The reverse is also true: tuning on a narrow task with clean data usually reduces hallucination. Which one you get depends on whether you are teaching behavior or facts.
- Inherited bias. A small training set carries whatever skew is in it, and the model adopts that skew consistently.
None of this makes fine-tuning a bad idea. A fine-tuned model is a specialist, and specialists are narrower by design.
Do You Need to Fine-Tune a Model Yourself?
This sounds like it needs a research team and a budget. It does not, and it costs less than you think.
Training is billed per million tokens of training data. In 2026 that runs from well under a dollar on open-weight providers to about twenty-five dollars on the priciest frontier route. A small classification project can land under a dollar; a large job, in the hundreds. Tuning a small open model with LoRA can cost under ten dollars of cloud time.
Beginners miss the second bill: some providers charge a markup to run a tuned model, others do not.
OpenAI announced in May 2026 that it is winding down its self-serve fine-tuning platform, closing new sign-ups during 2026 and ending new fine-tuning jobs for existing customers in January 2027, with the company pointing to newer models and prompt-based approaches covering more of what fine-tuning once did. Other routes stay open, often cheaper: hosted providers that tune open-weight models, managed tuning from other major providers, and tuning open models yourself with LoRA. The toolbox is not closing, it is moving.
The honest answer is probably no. Knowing the term still pays off: it explains why one AI product answers your industry’s questions better than another, and whether a vendor’s “custom model” claim means anything. With a narrow task and a few hundred clean examples, this is now a weekend project.
Fine-Tuning FAQ
What is the difference between fine-tuning and training a model from scratch?
Training from scratch builds a model’s general ability from nothing, which takes enormous amounts of data and compute and is done by a handful of large labs. Fine-tuning starts from one of those finished models and adds a short, cheap extra round of training on a small focused dataset.
How much data do you need to fine-tune a model?
Far less than most people expect. A few hundred to about a thousand clean, consistent examples is a normal starting point for a narrow task, and some techniques work with as few as a couple of hundred. Consistency matters more than volume.
Does fine-tuning stop an AI from making things up?
Sometimes, and sometimes the opposite. Tuning on a narrow task with clean data tends to reduce made-up answers, but tuning that pushes new facts into a model can leave it confidently wrong about things it previously had right.
Is a fine-tuned model better than the original?
Better at the one job it was tuned for, and usually worse at everything else. That narrowing is the point, which is why a fine-tuned model is a poor choice when you need one model to handle many different kinds of request.
