Fine-tuning

A large language model such as GPT-4 or Claude is trained on general internet data and can do almost anything. For specific applications, like assessing requests in your industry or writing in your company's fixed tone of voice, a general-purpose model is not optimal. Fine-tuning solves this by training the model on examples that show exactly what you expect.

How it works: you supply a set of examples in which the input and the desired output are fixed, for instance a customer question with the answer an experienced colleague would give. The model learns the pattern in those examples and applies it to new input. In practice this usually happens with parameter-efficient techniques such as LoRA, which adjust only a small part of the model. That keeps a project affordable and makes it easier to move to a newer base model later.

Fine-tuning is not a knowledge base. The model gets better at a fixed task, style or classification, but it does not learn new facts: information that changes after the project is not in the model. For current or changing knowledge you use RAG, where the agent consults your own sources for every question. The order that works in practice is: prompting first, then a connection to your own sources when the knowledge lives in documents or systems, and fine-tuning only when the model must apply the same style or the same assessment pattern consistently and prompts do not achieve that reliably.

What you need: a set of hundreds to a few thousand good examples, built consistently, with a clear line between what is acceptable and what is not. Quality matters more than volume. Record per example what the input was, what the desired output is, and why a variant was rejected. Without those rejections the model only learns what is right, not where the boundary lies.

A four-week project: in week 1 we define the task, the limits of what the model may and may not do, and which examples are usable. In week 2 you supply the examples and we set aside a set that is not used for training, so we can later measure whether the model really performs better. In week 3 the model runs on that separate set and a colleague reviews the outcomes, including the cases where the model hesitates. In week 4 the model goes into use within the agreed task, with human review at the moments you designate.

Where it goes wrong: too few examples, examples that contradict each other, or a task that in fact needs current information. A model fine-tuned on a hundred examples will not handle an industry in which every client is different. You also have to plan maintenance: a new base model or a changed way of working means repeating the training. Budget for a fixed evaluation set and an owner on your side who guards quality.

The cost is mainly in preparation, not in the compute time of the training: collecting and reviewing examples takes hours from your own team. At Match-AI an AI colleague starts with a 4-week onboarding period for €999 for up to 3 people or €2,495 for up to 10 people per 4 weeks, excl. VAT, and we only start fine-tuning if the initial analysis shows that prompting and a connection to your own sources do not solve the problem.