Is your data ready for fine-tuning?

How we read your answers

How many examples do you need?

Less than most people expect to start, more than most people expect to finish. We always begin with fifty real examples of the task with answers someone who does the work has checked. We freeze them as a test set that nobody trains on. On that set we measure untuned models first. If one of them already does the job, you do not need training, and you found out for the price of one day.

For a narrow task with a strict format, such as pulling six fields from an invoice or sorting tickets into your categories, fifty examples are also enough for a first trial model. That is how our one-week trial works. Broader tasks, free-form answers or a new language need more. How many more is something the measurement on your fifty tells you, not a rule of thumb.

The ceiling is set by quality, not quantity. For our Slovenian medical research model we first built 140,063 instruction records and passed every one through a quality gate before any training started. A thousand examples with unchecked answers teach the model your mistakes.

PDFs are source material, not training data

"Can we just fine-tune on our PDFs?" is the second most common question. Not directly. A model learns from pairs: this input, this answer. A PDF is neither. You have two options:

• Retrieval (RAG): the PDFs are indexed and the relevant passages go into the prompt. Good when the answer is a fact from the document.
• Training: the text is extracted (OCR for scans, layout for tables) and turned into pairs, for example a document and the JSON it should yield. Good when the model has to learn a format or a way of working.

Tickets and e-mails are closer to training data: the customer's message and the reply your team sent are already a pair. They still need cleaning, removal of personal data and a check that the replies were good ones.

Four possible results

• RAG. Content that changes weekly, answers that need a source, access that differs per document. Training does not help here. See Fine-tuning or RAG for the full comparison.
• SFT. Behaviour: a strict output format, your terminology, a smaller language, tool calls. In our public tool-calling fine-tune, correct "no call" decisions went from 0.847 to 0.925 on 2,471 held-out steps. Retrieval cannot give you that.
• CPT + SFT. The language or domain is new to general models and you have a large archive of text. The model first reads the archive (continued pre-training), then learns the task. After continued pre-training on 1.78 billion Slovenian tokens, perplexity of our 4B model dropped by 51 %. This is program-sized work.
• Not yet. No examples with correct answers. Neither training nor a fair comparison is possible. Collect fifty first.

Model cards and results for the numbers above are public on Hugging Face.

When you do not need us

• You send a few hundred requests a month to a large API, the answers are good, and the data may leave the company. Keep the API. The calculator on the price list shows when your own model starts to pay off.
• Nobody in the company can say what a correct answer is. Then no one can check the model either, ours or anyone else's.
• You need a guarantee for a clinical or legal decision. A tuned model does not give one. It can prepare a draft for a person who decides.

What the next step costs

If the check says "not yet", the next step is free: collect fifty examples. Otherwise it is one of two:

• Evaluation, €1,900, one day. Up to five models on a frozen set from your examples, a results table and a written recommendation on whether to train and what. The fee is credited against a pilot. Sample report.
• Pilot S, €3,900, 2–3 weeks. When you already have checked examples and a test set: one fine-tuned model, a base-vs-tuned scoreboard, a GGUF build and a go/no-go recommendation.

If we have to build the data, that is Pilot M (€12,000, 4–5 weeks). Programs with continued pre-training start at €20,000. All prices in EUR, excluding VAT, on the price list.