Skip to content

AI models and fine-tuning

When the generic model does not reach the accuracy you need

We tune models with your data and your vocabulary for the tasks where getting it right 80 % of the time is not good enough.

  • Your vocabulary and your criteria
  • Measured before and after
  • Can run on your own infrastructure

Fine-tuning is not the first answer to almost any problem. Before it, you exhaust good prompt design and anchoring to your documents, which are cheaper and faster to change.

But some tasks need it: classification with very specific criteria, extraction from documents with an unusual layout, technical vocabulary a generic model does not handle.

Part of our job is telling you which of the two situations you are in.

What we offer

The work, broken down

  1. 01

    Prior diagnosis

    We check whether fine-tuning is the answer or whether your problem is solved more cheaply another way.

  2. 02

    Dataset preparation

    The part that decides the outcome: selecting, cleaning and labelling the examples the model is tuned on.

  3. 03

    Model tuning

    Training on open or provider models, depending on your case and your data restrictions.

  4. 04

    Measured evaluation

    We compare before and after against a test set. If it does not improve enough, we say so.

  5. 05

    Deployment and upkeep

    Going live, monitoring quality and retuning when the data shifts.

How we work

The way we do it

  1. 01

    Data decides

    A good set of a thousand examples beats a bad set of a hundred thousand. Most of the effort goes there.

  2. 02

    Everything is measured

    Without a metric before and after, there is no way to know whether the tuning achieved anything.

  3. 03

    The weights are yours

    When we work with open models, the tuned model stays with you.

Are you sure you need fine-tuning?

Half the time, you do not. We tell you before charging you for it.

Have us check

What we work with

  • Language models

    We work with the main providers and with open models. We pick per case, not per contract.

  • Digitalisation Agent

    Accredited under Kit Digital, with calls currently open: part of the project can be delivered against your voucher.

  • Open technology

    Python, PostgreSQL, Docker. Nothing proprietary that locks you in.

  • Documented code

    Everything we build ships with its documentation and its repository. It is yours.

  • Noroeste Digital

    Our own brand for hosting, domains and professional email.

  • On-site work

    We work remotely across Spain and come to your premises when the project calls for it.

Frequently asked questions

When is fine-tuning worth it?

When the task is repetitive, has very specific criteria and high volume, and you have already exhausted good prompt design and document anchoring. Before that, almost never.

How much data is needed?

It depends on the task, but hundreds or a few thousand well-chosen examples usually suffice. Quality matters more than quantity.

Can it run on our own infrastructure?

Yes, working with open models. That is the usual route when the data cannot leave the company.

What happens when our data changes?

Quality is monitored and the model is retuned periodically. A tuned model is not a finished product, it is a system you maintain.

Shall we look at it with your processes on the table?

Tell us how you work today and we will tell you what is worth automating and what is not worth touching.