Custom LLM Fine-Tuning

Fine-tuning trains a model on your own examples so it behaves in a specific way by default — your writing voice, your classification categories, your domain's conventions. It is the right answer considerably less often than clients expect, and most of what people want from it can be achieved more cheaply another way.

Get a quote WhatsApp us

Who it's for

Businesses with genuinely specialised language or a large proprietary dataset.

What changes

A model that speaks your business's language.

Starting at: ₹75,000
Timeline: 2–4 weeks for an audit
Category: AI Consulting & Enablement
Built from: Vashi, Navi Mumbai

Key takeaways

Most problems clients bring to fine-tuning are solved better by prompting or retrieval.
Fine-tuning teaches behaviour and format, not facts — it will not make a model know your data.
It needs 500–5,000 quality examples; fewer produces an expensive non-improvement.
Every base model change means retraining, which is an ongoing commitment.
Projects run 6–12 weeks from ₹3,00,000, after a paid feasibility assessment.

The three approaches, and choosing between them

Almost every request for fine-tuning is really a request for one of three things, and picking the wrong one is expensive.

Need	Right approach	Why
Model should know our documents	Retrieval (RAG)	Fine-tuning does not add facts reliably
Model should follow our format	Prompting with examples	Usually sufficient, no training cost
Model should write in our voice	Fine-tuning	Genuine use case
Model should classify our categories	Fine-tuning	Strong use case with enough data
Model should use our jargon correctly	Fine-tuning or prompting	Depends on volume of terms
Model should be cheaper at scale	Fine-tuning a small model	Legitimate cost optimisation
Model should be more accurate generally	Neither	Change the model or the task

The misunderstanding that costs the most

Clients frequently want to fine-tune a model on their company documents so it "knows" their business. This does not work the way it sounds.

Fine-tuning adjusts how a model responds, not what it can recall. Train it on a thousand of your documents and it will learn to sound like your documents while still being unable to reliably state what is in any particular one. Ask it a specific factual question and it will produce something in the right style that may be wrong.

Retrieval solves that properly — the model reads the relevant document at question time and cites it. When a client wants factual accuracy over their own material, we build retrieval and say so, even though it is a smaller sale than a training project.

What good training data looks like

The examples are the project. A well-chosen thousand outperforms a poorly-chosen ten thousand, consistently.

They need to be genuinely representative of what the model will face, including the awkward cases rather than only the clean ones. They need to be consistent — if two examples handle the same situation differently, the model learns inconsistency. And they need to be correct, since errors in training data become errors in behaviour.

Preparing them is usually the largest part of the engagement and it needs your domain people, not ours. We build the process and the tooling; the judgement about what a correct output looks like is yours.

The ongoing commitment nobody mentions at the start

A fine-tuned model is tied to the base model it was trained on. When the provider releases a better base model — which happens every few months — your tuned model does not automatically benefit.

You then choose: stay on an ageing base and fall behind what the general models can do, or retrain, which costs money and time each cycle.

This is a real ongoing obligation and it should be part of the decision. For a use case with genuine, sustained value it is worth it. For a marginal improvement it is a recurring cost attached to a benefit that was never large.

When it is genuinely the right answer

Classification at volume, where you have thousands of labelled examples and need consistency a prompt cannot deliver — routing support tickets into your own categories, grading documents against your own criteria.

Voice and format, where output must match a specific style closely and repeatedly. A firm producing hundreds of documents in a distinctive house style is a real case.

Cost reduction at scale. Fine-tuning a small model to match a large model's performance on one narrow task can cut per-request cost substantially, and at high volume that pays for the training quickly.

Domain language, where your field's terminology is used differently from general usage often enough that prompting cannot cover it.

Engagement, cost and delivery

The feasibility assessment is ₹75,000 over two weeks. If it recommends proceeding, the training project runs six to twelve weeks from ₹3,00,000, including data preparation tooling, training, evaluation against a held-out set, and deployment.

You receive the trained model weights or the provider-hosted tuned model under your account, the training dataset, the evaluation framework, and the pipeline to retrain when you choose to.

The evaluation framework matters most in the long run. It is how you decide whether the next base model plus a prompt has caught up with your tuned model, which it sometimes has.

FAQ

Custom LLM Fine-Tuning — your questions

How many examples do we actually need?

For a narrow classification task, 500 to 1,000 well-chosen examples often suffice. For style and voice, 1,000 to 3,000. For complex behaviour across varied inputs, 5,000 upwards. Below a few hundred, fine-tuning generally produces no measurable improvement over good prompting and occasionally makes things worse by overfitting to a small set. If you do not have the examples, the first phase is generating them properly rather than proceeding with what you have.

Can we fine-tune an open model and run it ourselves?

Yes, and it is worth considering where data cannot leave your infrastructure or where volume is high enough that per-request API costs dominate. The trade-off is operational: you take on GPU hosting, scaling and maintenance. For most SMBs that is a real burden and a hosted tuned model is the better choice. For clients with regulatory constraints or genuine scale, self-hosting is the right answer and we build for it.

Will fine-tuning stop the model making things up?

No, and this is a common hope. Fine-tuning shapes style and behaviour; it does not install a factual grounding. A tuned model will still produce confident wrong answers, now in your house style, which is arguably worse because it reads more authoritatively. Reducing fabrication is a retrieval and verification problem — ground answers in retrieved sources and check claims against data. We would build that rather than tune for it.

How do we know it actually improved?

By evaluating against a held-out set the model never saw during training, using metrics defined before the work starts. This is set up as part of the feasibility assessment so the success criteria are agreed in advance rather than argued about afterwards. We report the tuned model against three baselines: the base model with a simple prompt, the base model with a well-engineered prompt, and the tuned model. If the second one is close, we say so.

What does it cost to run afterwards?

Hosted tuned models typically cost somewhat more per request than the equivalent base model, though a tuned small model is usually far cheaper than an untuned large one for the same task — which is often the point. Self-hosted models cost whatever the GPU costs, which is a fixed monthly figure rather than per-request. We model both against your expected volume during the assessment so the decision is made on numbers.

Want a Custom LLM Fine-Tuning for your business?

Tell us what the process looks like today and we'll tell you what it would look like automated — and what it would cost.

Request a quote Talk to us