Tone, format, the rules of your domain and three examples, sent again with every call. The model still misses your vocabulary, and your archive of good answers has never been shown to it.
Tone, format, the twenty rules of your domain, three examples. All of it is sent again with every request, and paid for every time.
Your product codes, your clause numbering, the way your analysts write a summary. A general model gets close and no closer.
Thousands of good answers sit in your archive. The model has never seen one of them.
Fine-tuning trains a model on pairs of input and the answer you want. The rules move from the prompt into the weights. The prompt gets short, the answers get closer to yours.
Three parts. The data comes first, and most of the work is there.
Good examples, cleaned, and enough of them.
The base model, the method, the runs.
The tuned model in front of your users.
Four phases, with a number at the end of each.
The task, the examples you have, and the gate.
Build the training set and the test set.
Runs, evaluation, the pick.
Serve, watch, hand over.
Shorter prompts, closer answers, and a model that is yours.
Instructions and examples move into the weights. Each call sends the task and little else, which costs less and runs faster.
Measured on the held-out set against the answers your team wrote. The number is reported before anyone relies on it.
Weights, training set and evaluation are handed over. You can retrain, move or retire it without us.
| Side by side | Prompting | Fine-tuned model |
|---|---|---|
| Your vocabulary | Explained in every prompt | Learned from the examples |
| Effort | Days to write, endless to maintain | Weeks to build, then retrain on a cadence |
| Accuracy on your task | Whatever the prompt reaches | Measured against your answers, gated |
| Tokens per call | Instructions plus examples plus task | The task |
| What you keep | A prompt | The weights, the data, the evaluation |
It depends on the task. A narrow task with a consistent format can show a measurable gain with a few hundred clean pairs. A broad task needs thousands. We count what you have in discovery and measure a prompted baseline first, so the decision to tune rests on a number.
Training runs on your infrastructure or on a route that does not retain data. The training set is anonymised before it is used, and the weights it produces are yours. Nothing from your examples reaches a shared model.
Your domain drifts: new products, new rules, new phrasing. We log inputs and outputs, and retrain on a cadence you set, typically when the live accuracy drops below the gate. Each version is kept with its training set so any release can be rolled back.
Yes. Open base models can be tuned and served inside your perimeter. We choose the base model in discovery with your hardware in mind, so the tuned model runs where the data is allowed to be.
We are a small team of senior specialists. We pick the right model and the right layer, and we build the least machinery that does the job. You get a call with an engineer, not a sales deck.
