Teach an AI Model Your Format, Labels and Style

01 What is LLM fine-tuning?

Training an existing large language model (LLM) further on your own examples, so it learns a task, a format or a style. With adapter methods such as LoRA, only a small set of added weights is trained and the original model stays unchanged. The result is a model that does your task without long instructions in every prompt.

02 Fine-tuning or RAG: which do we need?

Use RAG when the answers depend on facts that change or must be cited, such as prices, policies or stock. With RAG, the model searches your documents at the moment of the question. Use fine-tuning when the model must learn a way of working, such as a format, labels or a style. Our RAG development page covers the first.

03 When is fine-tuning not worth it?

When good prompts already do the job, which is often; OpenAI's own guide says prompt engineering may be all you need. Also when your facts change often, since RAG keeps them current, or when you have only a handful of examples. We measure prompting and RAG first, so you only pay to fine-tune when it helps.

04 Can you fine-tune Llama and other open models?

Yes. We fine-tune open models such as Llama, Qwen, Gemma and Mistral, then run the tuned model on your own servers or in your cloud. Open models also mean you can keep the tuned model files (the weights) and move them, within the base model's licence.

05 How many examples do we need?

It depends on the task, and quality matters more than quantity: consistent, correct examples beat a large messy set. In the pilot we hold some of your examples back to test the model, so the results show whether you have enough before you invest in more.

06 Does our training data stay private?

If you choose an open model on your own servers, the training data and the tuned model stay on machines you control. If you fine-tune a hosted model, such as one from OpenAI, the provider processes the data. We help you choose based on your rules.

07 Can you fine-tune OpenAI models too?

Yes, where OpenAI offers fine-tuning for the model you need. It can let a smaller, cheaper model handle a task a larger one did. The data is then processed by OpenAI, and the model runs on its service, so it is a choice for data that is allowed to go there.

08 Will a fine-tuned model know our latest prices or products?

No. A model learns from the examples it was trained on, and does not see later changes unless you retrain it. For facts that change, such as prices, stock or policies, RAG is the right tool, and the two can work together: a tuned model answering from documents that RAG finds.

09 How do you know the tuned model is better?

We test it on the test set, examples it never trained on, and score it next to the base model with good prompts, and next to RAG where it applies. It ships only if it wins on the measure we agreed at the start. If it does not, we tell you and you keep the simpler option.

10 How much does LLM fine-tuning cost?

Three things drive the cost: how ready your examples are, the size of the model, and where it will run. Tell us the task and what examples you have, and within two working days we come back with an approach, a rough timeline and a rough number.

11 How long does it take?

It depends on how ready your examples are and how many rounds of training the task needs. We start by setting the score to beat and running a pilot fine-tune, so you see the comparison before a full project. We agree the timeline at the discovery call.

12 Who owns the fine-tuned model?

You do. The tuned weights, the training code, the data steps and the documentation go to you, in your own code repository and on your machines, within the licence of the base model you chose. You can retrain or run it yourselves.

13 Do you work with UK and US clients from India?

Yes. We sign an NDA before we start, on request, and work only in the accounts you invite us into. The code goes into your own code repository, in an account in your company's name.