IT Glossary
What is fine-tuning?
Fine-tuning means training an existing AI model further on your own examples, so that it specialises in your company's tasks, format and tone of voice.
There are three ways to make a general AI model behave like it works for you, and they are not interchangeable. You can write better instructions in the prompt, which is free and instant. You can retrieve your own documents and hand them to the model as context, which is how most business assistants answer questions about company knowledge. Or you can fine-tune: take a pre-trained model and continue its training on a curated set of your own input-output examples, permanently adjusting its weights so the behaviour you want becomes its default rather than something you have to ask for each time. Fine-tuning is the right tool for form rather than fact. It teaches consistent output structure, domain vocabulary, a house style, a classification scheme peculiar to your business — things that are tedious to specify in every prompt. It is the wrong tool for knowledge that changes, because retraining is how you would update it. The commercial attraction is usually cost: a small fine-tuned model that matches a large general one on your narrow task can cut the per-request bill substantially, which matters at volume rather than in a pilot.
Let’s talk about your project
Message us on WhatsApp or send an email — you talk directly to a developer.
office@northdan.com · +40 752 070 247
Why it matters for your business
Specialisation in your own domain
The model absorbs your vocabulary, categories and output format, so results stop needing long instructions and manual correction.
Smaller models, smaller invoices
A tuned compact model often matches a much larger general one on a narrow task, which changes the economics at production volume.
Consistent brand tone and behaviour
Style, register and refusal rules become the model's default rather than something every prompt has to re-establish.
Frequently asked questions
When should I choose fine-tuning and when retrieval-augmented generation?
Ask whether the gap is behaviour or knowledge. If the model does not know your prices, policies or product details, retrieval is the answer — documents change and retrieval reflects the change immediately. If the model knows enough but formats badly, uses the wrong register or misclassifies your categories, fine-tuning fixes it. Many production systems use retrieval for facts and a light fine-tune for form.
How much data do I need for a useful fine-tune?
Less than people expect, and quality dominates quantity. Several hundred well-chosen, consistent examples often produce a visible improvement for a narrow task; a few thousand is comfortable. What ruins a fine-tune is inconsistency — two examples that answer the same input differently teach the model that both are acceptable, which is exactly the ambiguity you were trying to remove.
Can I fine-tune without exposing confidential data?
Yes, with attention to where training happens. Open-weight models can be tuned on infrastructure you control or in an EU region, so nothing leaves your perimeter. With hosted providers, check the contract for whether training data is retained or reused, choose the EU data residency option, and remove or pseudonymise personal data before training — a model can reproduce fragments of what it was trained on.
Let’s talk about your project
Message us on WhatsApp or send an email — you talk directly to a developer.
office@northdan.com · +40 752 070 247