- Home
- AI development
- LLM fine-tuning
LLM fine-tuning
Fine-tuning a language model on your examples, after the cheaper fixes are tested
Fine-tuning trains an existing model on a few hundred to a few thousand of your own examples so it writes in your format, tone or routine. Most businesses do not need it: better instructions or search over your documents fix most problems for less. We try those first, measure the result, and fine-tune only when they fall short.
- Prices published on the site
- First working version in two weeks
- Support month to month, cancel any time
Fine-tuning shapes how a model writes; current facts come from search
A fine-tuned model has been trained further on examples of the output you want: pairs of an input and the ideal answer. Afterward it produces your format without a long prompt. A claims summary in your template, a ticket tagged with one of your 40 categories, a reply in the voice of your best support person.
Fine-tuning does not reliably add knowledge. A price list or a policy changes every month, and a tuned model keeps answering from the old version until it is trained again. Facts like that belong in private AI over your documents, which looks up the current page at question time and cites it. We start with instructions and retrieval because they are faster to build and cheaper to change.
Instructions, retrieval and fine-tuning compared
| Compared on | Better instructions (prompting) | Retrieval over your documents | Fine-tuning |
|---|---|---|---|
| What it changes | What the model is told to do and how | Which facts the model can see when it answers | How the model writes and behaves by default |
| Data you need | A written task description and a handful of examples | Your documents: policies, manuals, past tickets | From about 50 to thousands of correct input and answer pairs, depending on the task |
| Time to a first result | Hours to days | Days to two weeks | Weeks, mostly spent collecting and checking examples |
| Cost to change later | Edit the text and rerun the tests | Add or replace documents, nothing to retrain | Train again on new examples and pay for GPU time again |
| Right for | Most first attempts: format, tone, rules for edge cases | Answers that depend on your files and change often | One narrow task at high volume, when the other two fall short |
We work left to right. Each step is scored on the same development set of your cases, and a separate set is kept back for the final check. We move to the next step only when the score shows a gap.
What a fine-tuning project includes
A test set from your cases
50 to 200 past cases with the answer you would accept, held back from training. The same set scores every version. See LLM evaluation.
A cleaned training set
Your past cases become input and answer pairs. Personal data that must not leave your systems is removed, and your team corrects wrong pairs.
A trained model
An open-weight model such as Llama or Qwen is trained with adapter training, often called LoRA, on rented GPU time. Where a hosted model fits better, we use a provider's fine-tuning service where one is offered for that model.
A before-and-after report
The base model and the tuned model scored on the same test set, plus the cases the tuned model still gets wrong.
When fine-tuning pays off, and when it does not
Worth fine-tuning when
- The same format or tone is needed on thousands of items a month, and a long prompt still drifts
- You have, or can build, a few hundred correct examples, such as past tickets with the category your team assigned
- The task is narrow and stable, and a smaller tuned model may cost less per request than a large one
Use something simpler when
- The model gets facts about your business wrong, which is a retrieval problem
- You have fewer than about 100 examples, or your people disagree on the right answer
- Your rules change every few months, so editing a prompt is faster than retraining
- Nobody has measured the current output yet, so nobody could tell whether tuning helped
Priced by the work, with GPU time billed to you
One workflow running end to end, or one internal tool.
Several systems made to work as one: integrations, roles, reports.
A fine-tuning project usually lands in connected systems: training data from more than one source, the tuned model, and the integration that calls it. A small job on one task with data that is already clean can fit one process. The test set and the instructions-and-retrieval baseline come first, and their result decides whether training goes ahead. We estimate the GPU or API bill before anything is rented.
Care $300 to $500 a month. Active $750 to $1,500 a month. AI model usage, GPU time and hardware are paid to the provider, on accounts in your name.
Questions about fine-tuning
Do we need to fine-tune a model?
Probably not. Better instructions fix most format and tone problems, and retrieval supplies current facts, and we test that it finds the right ones. We try both, score the results on your test set, and propose fine-tuning only if the numbers show a gap that training can close.
How many examples do we need?
It depends on the task: a fixed format can improve with 50 to 100 good examples, and a varied task needs hundreds or more. We measure as we add them. The examples have to be correct and consistent: if two of your people answer the same case differently, the model learns the disagreement. We hold back 50 to 200 cases that it never trains on, so the final test is honest.
Can fine-tuning teach the model our products and policies?
Not reliably. A tuned model can still state a price wrongly and cannot show where it read it. Products, prices and policies change, so they belong in retrieval, which looks up the current document and cites it. Some projects use both: retrieval for the facts and a tuned model for how the answer is written.
Who owns the tuned model?
With an open-weight model such as Llama or Qwen, the trained weights sit in your account and can run on your own server, see local AI setup. The base model's license still applies, and we check it before training. A model tuned through a provider's API stays on that provider's platform under its terms, and we tell you that before you choose.
What does it cost to run and to retrain?
GPU time or API fees are billed by the provider to your account, and the amount depends on the model size and the number of examples. When a provider retires the base model, the tuning may have to be redone on a successor, if the provider offers tuning for it. We keep the training data and the test set in your repository so that is a rerun, and MLOps covers who watches for it.
How do you know the tuned model is better?
We score the base model and the tuned model on the same test set of your cases, before and after, and give you both results. If tuning does not improve the score, we say so and you keep the cheaper option.
Related services
- LLM evaluation Test which model does your job best
- Private AI over your files Answers from your own files, on your server
- Local AI setup AI on hardware you own, set up on-site
- MLOps Run, monitor and update models in production
- Generative AI Drafts, summaries and documents from your data
- AI development Every AI service we build, on one page
Tell us what's slowing you down.
You get a free written review by email within two business days.