Fine-tuning
Training a model on your own data
on hardware you own
Fine-tuning means teaching an open model your own material so it writes and sorts the way your business does. I run it on my own machine. Most businesses do not need it.
Fine-tuning means taking a model that already understands language and training it further on your own material, so it answers in your words and your format. It runs on a machine you own. For most businesses it is also the wrong place to start.
The honest answer comes first. Most of what companies want from this is done by retrieval, which means the model reads your own documents when a question is asked instead of memorising them. Retrieval is quicker to build, easy to correct when a document changes, and it can show you which file an answer came from. Fine-tuning earns its place in a narrow set of jobs, and the table below says which.
No client has bought this yet
I run fine-tuning on my own hardware, on my own material, for my own work. No client has bought fine-tuning from me yet. You should hear that from me rather than find it out later.
The training runs in your building
It happens on hardware you own, inside your own network, and the finished model stays on that machine. Nothing is uploaded to an outside AI service to be trained on.
One narrow internal job at a time
It suits one repeated job done the same way every time: sorting your paperwork into your own categories, pulling fixed fields out of your own forms, writing in a house format.
Retrieval first, almost always
If what you want is to ask your own documents and get an answer with its source, you want retrieval. Read fine-tuning or retrieval before you spend anything on training.
The model is yours at the end
The trained model, the material it was trained on and the machine all stay with you, with the scripts and a written note on how to train it again.
What fine-tuning changes, and what it does not
A model arrives already knowing how to read and write. Fine-tuning adjusts it with examples of your work until its habits match yours: your categories, your tone, your format, the shape of your answers. It is good at making behaviour consistent.
What it does not do is make facts reliable. A fine-tuned model does not know your price list changed yesterday, and it cannot tell you which document an answer came from. Retrieval does both, because the document is read at the moment the question is asked. A company that asks me to train a model on its data is usually describing a retrieval job in fine-tuning words.
When it is worth doing, and when it is not
This is the table I draw on the call. The left column is what people ask for. The right is what I would actually build.
| What you want | What I would build |
|---|---|
| Answers from our policies, with the source shown | Retrieval over your own documents |
| Sort incoming paperwork into our own categories | Fine-tuning, on your past decisions |
| One assistant the whole office can ask about anything | Retrieval, on a server in your office |
| Everything written in our house format, every time | Fine-tuning, once that format is written down |
| Answers that change as prices and documents change | Retrieval, because the files move too often |
| A model that works with no internet, on our own site | Either one, on hardware you own |
What this needs from you
Material, a machine and a person who can judge the output. The material is examples of the job already done correctly and in a consistent way, which is the part most companies underestimate. A pile of mixed files is not training data until somebody has decided which of them is right.
The machine is a bigger one than the rest of my work needs, with a graphics card, bought in your name and kept in your office. A private AI server for records and Agents does not need that. Training does. And somebody in your team has to be able to look at two answers and say which one is correct, because without that there is nothing to measure against.
How this work is priced
This is the most expensive thing I build, and the machine is usually the largest part of it. It is quoted once, per project, after the audit, and the quote separates the hardware from the work so you can see which is which. The hardware is bought in your name.
What moves the figure is how much of your material has to be cleaned up before it can be used, how many jobs the model has to do, whether you already own a machine capable of training, and how hard it is to agree what a correct answer looks like. If retrieval would do the job, I will quote that instead, and it comes to less. There is no monthly software subscription either way.
How it runs
- 01
An honest check first
We look at the job you want done and decide whether retrieval would do it. Often it would, and then I build that instead and say so plainly.
- 02
Prepare the material
Your examples are gathered, cleaned and checked by somebody in your team who knows what a correct answer looks like. This is the slow part.
- 03
Train and compare
The model is trained on your machine, then compared against the plain model and against retrieval on the same questions, in front of you.
- 04
Handover
The model, the material, the scripts and the machine stay with you, with a written note on how to train it again when your work changes.
Built, not described
Questions people ask
How much does it cost to train a model on our own data?
It is quoted once, per project, after the audit, and it asks more of your hardware than anything else I build. What moves the figure is how much material has to be cleaned up, how many jobs the model must do, and whether you already own a machine that can train.
How much of our data do you need to fine-tune a model?
Enough examples of the job done correctly and done the same way each time. I cannot put a figure on it before seeing your material, because a small, clean, consistent set is worth more than a large pile of mixed files nobody has checked.
Do we need fine-tuning or RAG?
Retrieval, in most cases. If your question is about facts that live in documents, retrieval is the right answer and fine-tuning would cost more without answering better. Fine-tuning or retrieval walks through both with the decision written out.
Have you done this for a client before?
No. I run fine-tuning on my own hardware for my own work. No client has bought fine-tuning from me yet. Everything else on this site is work I have built and handed over. If you need a client reference for a fine-tuning project, I do not have one.
Does the model learn our data and leak it somewhere?
The training runs on your machine and the finished model stays there, so nothing goes to an outside AI service. A trained model can repeat material it was trained on, which is why what goes into the training set is chosen deliberately and written down.
Can you fine-tune a model without a graphics card?
Training needs a machine with a graphics card. Retrieval over your own documents does not, which is one more reason to try that first. The audit decides which machine the work needs, and you buy it in your name.
What happens to the model if we stop working with you?
It keeps running. The model file, the training material, the scripts and the machine are all yours, and I hand over the documentation with them. Another engineer can retrain it without me.
Next step