A language model that knows your jargon, codes, and house style

You have tested a language model, and it almost works. It does not know your jargon, mixes up product codes, and writes in a tone that does not fit your company. You work around it with ever longer instructions, and every new exception adds another paragraph. Meanwhile, your archive holds thousands of examples of how it should be done: reviewed cases, approved quotes, answered tickets. Those examples are exactly the study material the model is missing.

When fine-tuning is and is not needed

Most of the time it is not needed, and we say that first. A good instruction plus access to your own documents solves most problems, and AI document search is a separate service for that. Fine-tuning only comes into play when the model has to learn something you cannot fit into an instruction. One example is a classification you have years of examples for, such as routing an incoming complaint to the right department. Another is a fixed format the model keeps getting slightly wrong, such as an inspection report or a quote line with your own article codes. A third is a smaller model that runs faster and costs less than the large one, but must do your task just as well.

Fine-tuning is not a way to store knowledge in a model that changes tomorrow. Prices, stock levels, and opening hours belong in a database the model consults, not in the model's weights. What you do train is behavior: how the model reads, sorts, and words things, and which mistakes it no longer makes.

Your own data as study material

You already have the data. An accounting firm has thousands of reviewed documents. A wholesaler has years of orders that someone translated into the right article code by hand. A care provider has reports that an experienced colleague has checked. We collect those examples, remove the unclear and duplicate ones, and strip out personal data the model does not need to learn the task. A few hundred clean examples are worth more than ten thousand messy ones.

As a base, we use an open model whose weights are freely available, and we continue training it. The result is a model file that belongs to you, just like the source code in Custom software. You also get the training script, so a year from now you can retrain with new examples without having to call us.

Training on a server in the EU

Training runs on a GPU server in the EU. Your files do not go to a US provider, and they do not train anyone else's model. Afterward, the training data and the model sit in a place you control. If you want to run the model in production, you can do that on your own hardware or through Private LLM hosting in the EU. We record which sources went into the training and why, because that is what the EU AI Act will ask of you. For the formal side of that file, there is AI governance and audits.

Measuring whether the model improved

Before we start, we set aside part of your examples. The model does not see them during training. Afterward, we give the same questions to the original model and the trained model, and count which one is right more often. That way, you know whether the training helped, instead of just feeling that it did. Within 10 business days, you have a first trained version and that count. If the new model is not better, we tell you right away and stop, or we go back to the instruction and the documents. If it is better, we build it into the screens your staff already works in through AI implementation.