AI Model Fine-tuning
AI model fine-tuning
We fine-tune a model when prompting and RAG genuinely cannot do the job — and tell you honestly when they can.
The problem
Fine-tuning is the most over-recommended technique in AI consulting. Most of the time, a well-written prompt or a RAG system over your documents does the job better and at a fraction of the cost. Fine-tuning makes sense in a narrow set of situations: when you need a very specific output format that prompting cannot reliably produce, when you need to run inference at very high volume and a smaller tuned model is significantly cheaper, or when you need a model to absorb a style or domain so deeply that it would take a prompt longer than the context window to describe. Outside those situations, fine-tuning is the expensive answer to a problem that has a cheaper solution.
You get a small, tuned model that does your specific task well and costs less to run than the big general one.
Built for: Teams with a specific, well-defined task and enough labelled examples to train on — and who have already found that prompting alone is not sufficient.
What we deliver
Honest feasibility first. Before we touch a dataset, we run your task through prompting and RAG and measure the results. Fine-tuning is only on the table if those approaches fall short on your specific evaluation.
Dataset preparation. Good fine-tuning data is hard to produce. We help you define the labelling criteria, review the examples for consistency, and reach a dataset size that will actually improve the model rather than overfit it.
Training and evaluation. We train against a held-out evaluation set defined before training starts, so the quality numbers reflect real generalisation rather than memorisation of the training data.
Cheap serving of the result. A fine-tuned model is usually smaller than the general model it started from. We deploy it on infrastructure sized for that smaller model so the per-request cost reflects the actual work being done.
More in Generative AI
Generative AI Development
Your product does in seconds what took a human twenty minutes.
Generative AI Consulting
You stop guessing and start with a short, honest list of what to build first.
Generative AI Integration
Your existing tools get smarter without a migration project.
Not sure which of these fits? See the whole generative ai practice, or read what we build for your industry.
Tell us what’s slow.
Describe the job eating your team’s day. We’ll tell you straight whether an agent is the right fix — and if it isn’t, we’ll say so.
Frequently asked questions
- How much training data do we need?
- It depends on the task. Simple classification tasks can work with a few hundred labelled examples. Tasks requiring nuanced stylistic output may need several thousand. We estimate the data requirement during feasibility, and if collecting enough examples is not realistic we tell you before you start.
- Will a fine-tuned model go out of date?
- Yes. If the world your model was trained on changes — new product lines, new regulations, new terminology — the fine-tuned weights do not update themselves. That is one reason RAG is often preferable: updating documents is simpler than retraining. We account for this in the recommendation.
- Can you fine-tune on our proprietary data without it leaving our environment?
- Yes. Open-weights models can be fine-tuned entirely within your own cloud infrastructure. Your training data does not need to go to any external service. We set this up when data confidentiality requires it.