# LLM fine-tuning on your own data and terms

> LLM fine-tuning on your own data: first check if prompting or retrieval is enough, then train, compare against the base model and keep control.

URL: https://webvijayi.com/services/llm-fine-tuning/

Last updated: 22 November 2026

## The short answer

LLM fine-tuning trains an existing language model further on examples from your business, so it follows your terms, format and tone without long prompts. Web Vijayi does this for companies that already run a language model at volume and keep seeing the same faults. For facts that change often, retrieval from your own documents is usually the better first step.

Before any training, we test whether better prompts alone fix the problem, as current guides from OpenAI and Google advise. When they don't, we prepare the training data with you, tune the model, and show on test cases it has never seen whether it does better than the original before anything goes live.

## Who is LLM fine-tuning for?

Medium and large companies that already use a language model at volume and keep seeing the same faults. A support assistant gets your product names wrong, a drafting tool ignores the house format, or a model that sorts claims or tickets puts the same awkward cases in the wrong pile every week. At high volume, each repeated fault means staff putting the same mistake right by hand, again and again.

It also suits product teams building AI into their own software who need a smaller model that is cheaper to run, and companies with strict rules on where data may go, because a tuned open model can run on servers you control.

## Are better prompts enough?

Most faults can be fixed without training anything. We write down with you what a good answer looks like, build a test set from real cases including the awkward ones, then design and test the instructions and examples the model receives. Every prompt version is stored with its test score, so nobody ships a change that fails the tests. Passwords, keys and access rules stay outside the prompt and are enforced by your own systems, as OWASP's security guidance for language models advises. If prompts reach the target, you stop here.

## What does LLM fine-tuning involve?

When prompts plateau, we train. Supervised fine-tuning teaches the model from pairs of inputs and correct answers, and suits clear tasks such as filling a form, sorting a message or writing in a set format. Preference tuning teaches it from pairs of better and worse answers ranked by your own experts, and suits work where "better" is a judgement, such as tone.

We usually train a small adapter on top of a frozen base model rather than changing the whole model, a method called LoRA (low-rank adaptation). The adapter is a separate file, so you can keep one per task and go back to the original model at any time. After each run we test again for safety, because a 2024 study presented at ICLR found that tuning on ordinary task data can weaken a model's safety behaviour.

## Gen AI model training for teams building their own model

Some clients are building a generative model or an AI product of their own. For them we prepare the data each stage of training needs: instruction and answer pairs, answers ranked by people for human feedback, examples from your field, and the test sets that measure progress. We agree the languages, subject areas and volumes with you before the work starts. For labelling at large volume, see [data annotation](https://webvijayi.com/services/data-annotation/) and [AI training data](https://webvijayi.com/services/ai-training-data/).

## Fine-tuning or your own documents?

Fine-tuning changes how a model works, meaning the terms it uses, the format it follows and the decisions it makes. It's a poor way to keep facts current, because every new price or policy would mean another training run. Facts that change belong in a document store the model reads when it answers, which is what [Ask your documents](https://webvijayi.com/services/ask-your-documents/) builds. A system can use both, with a tuned model reading from your latest documents.

## How it works

1. **We agree what "better" means.** Together we write success measures as tests and build a test set from real cases. Part of it is kept back and never used for training, so the final comparison is fair.
2. **We try prompts and record the score.** That score is the baseline every later step has to beat.
3. **We prepare the data with you.** You choose the sources. We clean them, take out personal details the task doesn't need, check each answer is supported by its example and have reviewers label to one written guide. You approve the dataset before training starts.
4. **We tune and compare.** The tuned model, the original and the prompt-only version all run on the same held-back tests, including safety tests. If you want deeper attack testing as well, we can [test the tuned model before launch](https://webvijayi.com/services/ai-evaluation-red-teaming/) with AI red teaming.
5. **You decide on go-live.** You see the results side by side and say yes or no. The previous version stays ready, so you can switch back.
6. **We keep it current.** When your products, rules or customers change, we add new examples and retune, and the same tests show whether the new version is better.

## What you get

- A written test plan with success measures and a held-back test set you keep
- Tested prompts with a version history and the score of each version
- A training dataset you approved, with a record of where each part came from
- The tuned model or adapter files, stored where you choose
- A results report comparing the tuned model with the original on the same tests, safety tests included
- Training records for every run: the base model, the data version and the settings
- A way back to the previous version at any time

## You stay in control

Everything we set up works with a person in charge. Nothing is trained on data you haven't approved, and the tuned model replaces nothing until you have seen its test results and said yes. Where we also [connect the model to your systems](https://webvijayi.com/services/ai-integration/), it can prepare answers and wait for approval, pass unusual cases to your team and keep a record of what it did. You can check it, change it, pause it or switch back to the previous version whenever you like.

While the project runs, you have a named contact at Web Vijayi and a written report at each stage. If your plans change, tell your contact and we change the plan with you or stop the work.

## What we need from you

- One person who knows the work and can say which answers are right
- Real examples, such as past questions and answers, tickets, reports or documents, with permission to use them for training
- Your rules: what the model must never say and which data must never be used
- Time from your experts to review samples and rank answers
- A decision on where the model must run, and any security review your company needs
- Your legal adviser's check that the data may be used for training and that the base model's licence suits your use

Fine-tuning usually doesn't make you a model provider under the EU AI Act. The European Commission's guidelines treat a company that modifies a model as its provider only if the change uses more than a third of the compute that trained the original, and ordinary fine-tuning usually stays well below that. Your duties for the system you build on it still apply, and your legal adviser should confirm the position for your case.

## Price

Every project is quoted in writing after a free consultation, with the timeline agreed in the same quote. Training compute and model hosting are paid separately. Full details on the [pricing page](https://webvijayi.com/pricing/).

## Next step

Tell us on a free consultation which answers your current model gets wrong. We'll tell you whether better prompts, your documents or fine-tuning is the fix.

[Book a free consultation](https://webvijayi.com/contact/#book) · [Get your free AI plan](https://webvijayi.com/free-ai-plan/)

## At a glance

- Paid separately: Training compute and model hosting
- Your data: Trained only on data you approve; Personal details the task doesn't need are taken out; A record of where each part of the data came from
- Next step: A free call, then a fixed quote in writing. You don't have to go ahead.

## Frequently asked questions

### What is the difference between prompt engineering and fine-tuning?

Prompt engineering changes the instructions and examples you send the model each time. Fine-tuning changes the model itself by training it further on your examples. Current guides from OpenAI and Google both say to get the most out of prompts first and to fine-tune only for errors that prompts cannot fix, so we always measure the prompt-only result before we suggest training.

### How much data does fine-tuning need?

It depends on the task, and quality matters more than volume. Microsoft's current guide suggests 50 to 100 good examples for a first test and 500 or more for production, and Google's speaks of hundreds to thousands. We size it properly once we have seen your examples and how often your own experts agree on the right answer.

### Should we use fine-tuning or RAG?

Fine-tuning changes how a model works by training it on your examples: your terms, formats, tone and decisions. RAG (retrieval-augmented generation) hands the model relevant passages from your documents each time it answers. Fine-tuning won't teach a model your latest prices and policies, because they change too often to retrain for each one, so they should stay in your own documents for the model to look up. Our Ask your documents service sets that up. Many systems use both, with a tuned model reading from your latest documents.

### Who owns the tuned model, and where does it run?

We agree ownership in writing before any work starts. If you need to hold the model yourself, we tune an open model whose licence suits your use and hand over the adapter files, the data and the training records. Hosted tuning services change: OpenAI's own guide, read in November 2026, says it is winding down its fine-tuning platform, and a model tuned on a hosted service lasts only as long as its base model.

### Is our data safe during training?

We use only the data you approve, take out personal details the task does not need, and record where every part of the dataset came from and which version was used. Training runs where you choose. Before we start, your legal adviser should confirm that the data may be used for training, because under data protection law reusing personal data for a new purpose needs its own check.

### How do you know the fine-tuning worked?

Before any training, we write down with you what a good answer looks like and build a test set of real cases with agreed answers. Part of it is kept back and never used for training. The original model, the prompt-only version and the tuned model are all scored on those held-back tests, safety tests included, and you see the results side by side before you decide on go-live.
