What is fine-tuning?

Fine-tuning: definition

Large language models are first trained in a general way, which makes them versatile but not tailored to your company. In fine-tuning, such a model is trained further with a few hundred to thousands of your own example pairs, each consisting of an input and the desired output. This changes its behavior permanently.

The distinction matters: fine-tuning mainly conveys style, structure and task routine, not reliably new factual knowledge. Current knowledge, such as prices or manuals, is better supplied at runtime through retrieval-augmented generation (RAG), because documents are easier to swap than a trained model.

How fine-tuning works

First, examples are collected and cleaned: good inputs with the correct output, preferably reviewed by experts. A portion is set aside for checking. Then the model is trained. Parameter-efficient methods such as LoRA are widespread; they train only a small add-on and therefore need less computing effort.

Afterwards, the result is measured on unseen test cases and compared with the original model. Only when it is measurably better does it go into operation. This prevents an improvement in one place from unnoticeably affecting another. If the provider changes its base model, fine-tuning usually has to be repeated.

Real-world example: consistent classifica­tion of incoming mail

A tax advisory firm receives letters every day from tax offices, banks and clients and wants to sort them into fixed categories, each with a deadline and a short summary in a set format. With instructions alone, the output fluctuates. After fine-tuning with several hundred reviewed examples, the format is stable and the classification more reliable.

The firm kept human control over deadlines. It also collected employees’ corrections so it could improve the model in a targeted way later. The example shows that fine-tuning pays off for a clearly defined, frequent task with many good examples. For changing content such as price lists, RAG would have been the better choice.

When it pays off—and when not

The usual sensible order is: prompting first, then RAG, fine-tuning last. It costs time and data, has to be maintained and can lock in errors if the examples are poor. If personal data is used for training, data protection and purpose limitation must be clarified. If you use examples from real cases, also check whether they contain confidential details.

Nano Concepts helps with the decision and, if needed, builds the data basis, the training and the quality checks. That includes a comparison with simpler solutions and a pilot that measures quality on real cases before a larger training run starts. We also say openly when fine-tuning is not needed.

• Define the goal and the metric in advance
• Compare with prompting and RAG before training
• Review examples professionally and remove personal data
• Keep test data separate and measure quality regularly
• Plan maintenance effort for model changes