A part of Coaley Peak®Fine-Tuning
The process of taking a pre-trained AI model and training it further on a smaller, specific dataset to improve its performance for a particular task.
Fine-tuning is the process of taking an AI model that has already been trained on a large, general dataset and continuing its training on a smaller, more specific collection of examples. The result is a model that retains its broad knowledge but has become noticeably better at a particular topic, style, or task. Think of it like a solicitor who studied law broadly at university and then spent several years specialising in employment law: the general training is still there, but a specific area has been deepened through focused practice.
This sits inside the wider topic of machine learning, and it is one of the most practical techniques for organisations that want an AI system aligned to their own context rather than a generic one. For background on how the base models are built in the first place, how LLMs work is a good starting point.
How it works#
A large language model starts with billions of parameters -- the numerical weights that encode everything it has learned during initial training. Fine-tuning runs a supervised learning process on your specific dataset, making small adjustments to those weights so the model's outputs align better with the new examples. Because the starting point is already a capable model, you need far less training data and far less computing power than building from scratch.
The training data for fine-tuning is typically a set of input-output pairs: a prompt and the ideal response. The model generates a response, compares it to the ideal, calculates the error, and adjusts its weights slightly to reduce that error. Repeat this across thousands of examples and the model shifts towards the target behaviour.
Fine-tuning is different from prompting, where you guide the model's behaviour in real time by writing careful instructions. With fine-tuning the adjustment is baked into the model itself, so you do not need to repeat detailed instructions with every request.
Examples#
Customer service for a specialist retailer. A cycling accessories retailer might fine-tune a language model on thousands of their past support conversations, product descriptions, and policy documents. The resulting model answers questions in the company's tone, knows their catalogue inside out, and handles returns queries without needing a prompt that explains all of that context every time.
Medical summarisation. A clinical software provider could fine-tune a model on de-identified patient notes and discharge summaries. The adapted model learns the conventions of clinical writing -- abbreviations, structure, terminology -- and produces summaries that match what clinicians actually write, rather than generic prose that a base model might produce.
Why it matters for business#
For most people using AI day-to-day, fine-tuning is invisible: it is something that happens before a product reaches them. But understanding it matters because it explains why some AI tools are dramatically better at specific tasks than others, and why a generic chatbot often falls short for specialist work.
Organisations considering AI adoption should be aware that fine-tuning is now an accessible option, not just something reserved for large technology companies. Cloud providers including OpenAI and Google offer fine-tuning as a managed service. The main investment is in preparing clean, representative training data, which is usually more time-consuming than the technical side.
To understand more about how AI is being applied in practical business settings, see the guide to AI for small business in the UK or explore the full learning paths.
Fine-Tuning: frequently asked questions
- How is fine-tuning different from just writing a better prompt?
- Prompting tells the model what to do in real time, but the model's underlying knowledge and style do not change. Fine-tuning actually updates the model's internal weights using new training examples, so the adapted behaviour becomes permanent and consistent without needing a lengthy prompt each time. For many specialist applications, fine-tuning produces noticeably more reliable results than even very detailed prompts.
- How much data do I need to fine-tune a model?
- It depends on the task and the starting model, but fine-tuning typically needs far less data than training from scratch -- often a few hundred to a few thousand high-quality examples are enough. The key is that the examples are relevant and accurate. Poor-quality training data will teach the model bad habits that can be hard to reverse.
- Is fine-tuning expensive?
- The cost has fallen considerably as cloud providers now offer fine-tuning as a managed service. For smaller models and modest datasets, a fine-tuning run can cost tens of pounds rather than thousands. Larger models and larger datasets cost more. Most providers charge per token of training data processed, so keeping your dataset lean and clean keeps costs manageable.
Go deeper with related tutorials
Useful? Now picture your whole team knowing it.
Everything here is free, and the first course is too: eight short online lessons your team can start this afternoon. When you want the skills to stick, we come to you for a day on-site, built around the work your people actually do.