AI glossary · Models
Fine-tuning
In one sentence
Fine-tuning means further training an AI model that's already been pre-trained, using a small set of your own examples, so it specialises in a particular task, format or style without starting from scratch.
Also known as: model fine-tuning · LLM fine-tuning · fine tuning · supervised fine-tuning
What is fine-tuning?
Training an LLM from scratch takes vast amounts of text and computing power, which only a handful of companies can afford. Fine-tuning starts from a model that can already write and gives it a second, short and focused round of training using examples of what you want it to do.
Put it to work
UTIO's AI, through an API
From €29/mo + VAT
For developers and agencies building their own products.
See the AI APIIt’s a bit like hiring someone experienced and showing them how things are done at your company: you don’t teach them to speak, you teach them your quote format or how you categorise support tickets.
Fine-tuning is good at changing behaviour: output format, tone, the way it classifies things. It’s poor at memorising data that changes. That’s what RAG is for.
Related to this term
AI API
- Transcription, voice, text, image, OCR and embeddings
- Per-client keys with a spending cap
- Prepaid wallet: €50 minimum top-up
From €29/mo + VAT Panel account, plus usage
See the AI APIHow it works
- Gather examples. Pairs of input and ideal output: “this email” → “this category”, “this data” → “this summary”. Usually hundreds or thousands, checked by hand.
- Train. The model processes the examples and adjusts its parameters to bring its answers closer to the ideal ones. Techniques such as LoRA adjust only a small fraction of the parameters, which makes the process much cheaper.
- Evaluate. Test it on examples it hasn’t seen, to check it generalises rather than parroting.
- Deploy the tuned model and keep an eye on it: if the task changes, you’ll need to tune it again.
Fine-tuning is widely used to specialise small models. A fine-tuned small language model can match a much larger model on one specific task.
Example
A lift maintenance company in Manchester receives fault reports by email and sorts them into twelve internal categories, each with a priority. A general-purpose model gets most of them right but mixes up categories only its engineers tell apart, such as “stopped at floor level” and “stopped between floors”.
With a few hundred reports the team has already categorised, fine-tuning teaches the model that distinction, and it then classifies the way the company does. Details about each building, such as the address or contract, aren’t trained in: they’re looked up separately.
Fine-tuning at UTIO
Most small businesses don’t need to fine-tune a model. UTIO’s agents adapt to your business through your knowledge base and the instructions you set, and when a price changes you simply update the document.
If you’re building your own solution with the AI API, you can choose the model tier for text generation and create embeddings to run your own RAG. Before you consider fine-tuning, try good instructions and good examples: they often go further than you’d expect, and they spare you from maintaining a model of your own inside your AI agent.
Frequently asked questions
Fine-tuning or RAG: which do I need?
If you want the model to know your data (prices, opening hours, catalogue), it's almost always RAG: it looks the data up for each question and you just update the document. If you want the model to change how it responds (a fixed format, a tone, a very specific classification), fine-tuning can help. The two are often combined.
How many examples do you need for fine-tuning?
It depends on the task and the model. Adjusting a format or style can take a few hundred good examples; harder tasks may need thousands. Quality and consistency matter more than volume.
Does fine-tuning stop a model making things up?
Not on its own. It can improve format and accuracy on the trained task, but the model still generates probable text. For answers grounded in specific facts, the effective approach is to pass the sources in with every query and tell it not to answer without them.
Related terms
- RAG
RAG (retrieval-augmented generation) is a technique in which a language model first retrieves the relevant passages from your own documents and then writes its answer from them, instead of relying on what it memorised in training.
- Model training
Model training is the process in which an AI model learns from a dataset: it's shown examples, its errors are measured and its parameters are adjusted, over and over, until its outputs are good.
- Dataset
A dataset is an organised collection of examples, such as text, audio, images or records, used to train, fine-tune or evaluate an AI model. Its quality directly shapes what the model learns.
- Small language model (SLM)
A small language model (SLM) is a language model with far fewer parameters than an LLM. It knows less about everything, but it replies faster, costs less and can run on a modest server, a laptop or even a phone.
- Model parameters
A model's parameters are its internal numeric values, mainly the weights of a neural network, which are adjusted during training and store what it has learned. Their count is used as a rough measure of model size.
Updated on