Skip to main content
Mobius
Advanced

What is fine-tuning

Also known as: fine tuning, model fine-tuning

Definition

Fine-tuning is extra training of a ready AI model on your own examples, so it learns a style, format or task. It changes how it answers, not what it knows.

Why it matters

Most owners who ask about fine-tuning need a clearer prompt and one current document instead, and knowing the difference saves a developer project.

Common mistakes

Fine-tuning to teach the model facts, then finding the answers unreliable and out of date after the first change.

Related terms

Worked exampleIllustration

Making a fictional pilates studio's assistant sound like the owner

A fictional neighbourhood pilates studio wants its WhatsApp assistant to sound like the owner. The steps and numbers are invented for this example.

  1. Tone or facts?

    It is tone. The studio's rules, such as the 24 hour cancellation, belong in the policy sheet the assistant reads, not in training.

  2. Prompt first

    Five tone rules and three sample replies go into the assistant's instructions, about an hour of work.

  3. Test

    The result is judged on 20 new customer messages.

    20 messages

  4. Only then

    Fine-tuning comes in only if the prompt still fails on a task repeated hundreds of times: about 100 pairs of a real question and the owner's reply, names and phone numbers removed, trained through the vendor's developer tools.

    about 100 pairs

Most small businesses stop at step 2. Fine-tuning changes how the model writes, and your facts stay in your documents. The 100 pairs are this example's assumption, not a rule.

Illustration

Prompt, RAG or fine-tuning?

  • Answer with our 24 hour cancellation rule

  • Write in our short, warm tone

  • Always answer in three short lines

  • Use the new rule the day the sheet changes

  • Keep the owner’s tone on hundreds of replies, after the prompt failed

Pick a tool for each need. Most small businesses stop at the prompt.

Facts go in the sheet (RAG). Tone and format go in the prompt first. Fine-tune only a repeated task the prompt already failed, with about 100 real pairs.

Three ways to shape an assistant. Pick the one that fits each need and see why.

Only the tone changes

CustomerCan I still cancel tomorrow’s class?

Base modelThank you for reaching out. According to our policy, cancellations must be made at least 24 hours before the scheduled class. Later cancellations count as a used class. Please let us know if you need any further assistance.

Facts from the sheet, the same in both

Up to 24 hours beforeLater counts as used

The base model gets the facts right and sounds like a form letter.

Fine-tuning changes how the model writes, not what is true. The facts come from the policy sheet in both replies.

Definition

Fine-tuning takes a model that has already been trained and trains it a little more on examples you provide: pairs of an input and the answer you want. The model's own settings change, so the new behaviour stays without being repeated in every prompt. It suits a consistent style, a fixed format or a narrow task repeated many times, such as sorting messages into types. It is an unreliable way to add facts and a slow one to update when a fact changes, which is why facts belong in documents the model reads with each question (RAG). Fine-tuning is a developer feature of the vendor's platform. Uploading files to a chat tool's project is not fine-tuning: the tool searches the files, and the model stays the same.

Why it matters

The prompt sets instructions and tone, RAG supplies your facts, and fine-tuning is worth it only for a narrow task repeated at volume where a good prompt still fails. There it can let a smaller model do one job well.

Directly related: RAG, Prompt, Model Training.

Improvement tips

  • Try a clearer prompt with three or four sample replies before you consider fine-tuning, and test it on 20 real messages.
  • Keep facts that change, such as hours and rules, in a document the model reads, not in training.
  • If you do fine-tune, collect dozens to hundreds of good examples, depending on the task, all in the style you want.
  • Remove names, phone numbers and ID numbers from customer conversations before they become training examples.

Common mistakes

  • Calling a chat project with uploaded files a fine-tuned model, when the model has not changed.
  • Training on inconsistent examples, so the model learns the mistakes along with the style.
  • Starting a fine-tuning project before a good prompt has been tried and measured.

Related terms

AI and DataAdvanced

RAG

RAG (retrieval-augmented generation) is a method where AI answers from your documents: it first finds the relevant passages, then writes the answer from them.

AI and DataBeginner

Prompt

A prompt is the instruction you give an AI tool: the question, the task, the context and the format you want. A clearer prompt gets a more useful answer.

AI and DataIntermediate

Model Training

Model training is how an AI model learns: it sees many examples, checks its guesses against the right answers and adjusts itself until the errors shrink.

AI and DataIntermediate

AI Model

An AI model is the trained file at the heart of an AI tool: it takes an input, applies patterns learned from data, and returns a prediction, a label or text.

AI and DataIntermediate

LLM

An LLM (large language model) is AI trained on masses of text to predict the next piece of a sentence. That is how chat tools write answers, right or wrong.

Quick check

A studio's opening hours change every season. Where should they go so its assistant answers correctly?

Choose an answer

Frequently asked questions

What is fine-tuning in simple words?
Extra training for a model that already works. You show it many examples of a question and the answer you want, and it adjusts itself to answer that way. It changes how the model answers, its style, format or a narrow task, rather than what it knows.
I uploaded my files to a custom chatbot or a project. Did I fine-tune it?
No. Those features combine your instructions with a search over the files you upload, and the model itself does not change. Anthropic's help page on Projects, for example, says Claude uses a project knowledge search tool to retrieve relevant information from your uploaded documents. Fine-tuning is a separate developer feature of the vendor's platform.
When is fine-tuning worth it for a small business?
Rarely, and only after the simpler steps. It fits a narrow task repeated hundreds of times, where a good prompt with examples still fails, such as sorting messages into fixed types or writing in one strict format. For answers about your business, a current document the model reads works better.
How many examples does fine-tuning need?
It depends on the task. OpenAI's guide, for example, sets a minimum of 10 examples, says it sees improvements from 50 to 100, and adds that the right number varies greatly. Plan on dozens to hundreds of good examples, and quality matters more than count.
Why did our fine-tuned model give an old answer?
Because what it learned in training stays until it is trained again. Fine-tuning is a poor place for facts that change, such as hours, rules or stock. Put those in a document the model reads with every question, and update the document instead.
Our prompt works on most messages but misses our tone. Should we fine-tune?
Not yet. Add three or four sample replies in your own voice and a short list of tone rules to the instructions, then test on 20 new messages. Fine-tuning comes in only if that still fails on a task repeated at volume.
Does fine-tuning stop the model from making things up?
No. It teaches style and format, and a tuned model can still invent a fact in the new style. To reduce invented answers, give it your documents to answer from and keep a person checking what reaches customers.
Who does the fine-tuning?
Usually someone technical, through the vendor's developer platform: preparing the examples, running the training and testing the result on new messages. The owner's part is supplying good examples and judging whether the answers sound right.
Is it safe to use customer conversations as training examples?
Only with care. Real conversations contain names, phone numbers and sometimes health or payment details. Remove them before the examples leave your hands, and check what the platform does with uploaded data. Israel's Privacy Protection Law, 1981, as amended by Amendment 13, in force since 14 August 2025, governs collecting, keeping and using personal details, including details you paste into or upload to an AI tool. Educational material, not legal or tax advice. Check your own case with a lawyer or an accountant.
Will fine-tuning make our assistant sound exactly like us?
It can get close on a narrow, repeated kind of reply if the examples are yours and consistent. Test it on new messages, not on the examples it learned from, and expect to review replies before they go out.
What is the difference between fine-tuning and RAG?
Fine-tuning changes the model's behaviour through extra training. RAG leaves the model as it is and gives it relevant passages from your documents with each question. Use RAG for facts, fine-tuning for a style or a narrow task, and the prompt for everything you can simply say.
Do I lose the original model when I fine-tune?
No. Fine-tuning creates a separate tuned version, and the original model stays available. If the tuned version disappoints, you go back to the original with a better prompt.

Sources: OpenAI, "Supervised fine-tuning" guide (developers.openai.com/api/docs/guides/supervised-fine-tuning), read 2026-09-25, OpenAI, "Model optimization" guide (developers.openai.com/api/docs/guides/model-optimization), read 2026-09-25, Anthropic Help Center, "Retrieval augmented generation (RAG) for projects" (support.claude.com/en/articles/11473015), read 2026-09-25, Privacy Protection Authority, Amendment 13 questions and answers (gov.il/he/pages/tikun13_qa), read 2026-08-31

Last reviewed: 2026-09-25

I have worked with

What Clients Say

From different fields, at different stages of business

Dan Manto

Eclipse Capital

Real Estate Investment and Finance, USA

Arty McLabin

GameReady

Game Development Education and Outsourcing, International

Anna, Beautician

Anna

Beautician

After I finished a cosmetology course at a leading company, I was confused. I didn't know how to open a business, what to sell from home, or how to bring in clients. Since I reached out to Mobius, everything has changed. We built an organized plan with clear steps and no unnecessary risks. Today I see results, and I'm growing and developing every month.

Mark, Massage therapist, Gan Yavne

Mark

Massage therapist, Gan Yavne

When I finished a massage course, I imagined a different world. I found out it's a hard field, and financial stability isn't easy to find. When I came to Mobius, something changed. We started an organized process and built a business plan. Without Mobius? I probably would have given up. Today I believe in myself and in my own path.

Wondering where AI actually fits your business?

Alex helps Israeli businesses use AI and data where it pays off, and skip it where it does not. The first call is free.

+972 055-248-6151. Better on WhatsApp: I am in meetings most of the day and answer as soon as I am free.