AI Model Fine-Tuning
Usually the wrong tool. Occasionally exactly the right one.
What is AI model Fine-Tuning?
Fine-tuning adjusts a model on examples of your own so it adopts a particular style, format or behaviour, as distinct from giving it access to your facts, which is a different technique entirely.
Fine-tuning is asked for far more often than it is needed. It sounds like the way to make a model know your business, and that is not what it does.
Fine-tuning teaches behaviour: tone, structure, how to respond to a kind of request. It does not reliably teach facts. If you want a model to know your prices or policies, retrieval is the correct approach and it is cheaper, faster and updateable.
Where fine-tuning genuinely earns its place is narrower. A consistent output format, a specialist tone, or a repetitive classification task where prompting has been tried and is not consistent enough.
Why we try the cheaper options first
Duara fine-tunes models from Nairobi, and turns down most fine-tuning enquiries because the client wants the model to know something rather than to behave differently.
We will not fine-tune before prompting and retrieval have been tried properly. Those two solve the large majority of cases at a fraction of the cost, and a fine-tune that could have been a prompt is money spent on something you then have to maintain.
Who this is for
- Prompting has been tried and is inconsistent
- You need a strict output format every time
- You have a repetitive classification task at volume
- You hold hundreds of good examples already
Industries we serve
- Software companies
- Financial services
- Legal and professional services
- Healthcare technology
- Media and publishing
- Customer operations at scale
Problems this solves
These are the narrow conditions where fine-tuning is the right answer rather than the expensive one.
- Prompting is inconsistent at volume. Instructions work most of the time and fail unpredictably, which matters when the output feeds another system.
- Output format drifts. A structured response is required every time and the model occasionally departs from it, breaking whatever consumes it.
- Prompts have grown enormous. Instructions and examples now consume most of the context and cost, which fine-tuning can compress.
- A specialist tone is needed. Generic model output does not match how your organisation writes, and prompting gets close without getting there.
What we deliver
An honest assessment of whether fine-tuning is the right tool, a dataset built and cleaned properly, an evaluation set held back before training, and measurement against the untuned baseline.
An assessment of whether fine-tuning is right at all
A training dataset built and cleaned from your examples
An evaluation set held back before training
Measurement against the untuned baseline
Deployment through your own API
A retraining plan for when base models change
Features
- Prompting and retrieval tried first. Those solve most cases at a fraction of the cost, and we will not skip them to reach the larger project.
- Evaluation set held back first. Test examples are separated before any training, because measuring on data the model saw proves nothing.
- Compared against the untuned model. The baseline is measured, since a fine-tune that performs no better than a good prompt is not a success.
- Dataset quality treated as the work. Most of the effort is building and cleaning examples, and a small clean set beats a large inconsistent one.
- Retraining planned for. Base models change and a fine-tune does not follow automatically, which is a maintenance commitment we state upfront.
Benefits
- Consistent behaviour at volume. Where prompting was nearly right, a fine-tune removes the unpredictable failures.
- Shorter prompts and lower cost. Behaviour learned by the model does not need repeating in every request.
- Reliable output structure. Downstream systems can depend on the format rather than defending against variation.
- You find out whether it was worth it. Baseline comparison tells you honestly, including when the answer is no.
Technologies we use
- Suitability assessment. Establishing whether the goal is behaviour or knowledge, which determines the technique entirely.
- Dataset construction. Building, cleaning and balancing examples, which is where most of the real work sits.
- Evaluation design. Holding back test examples before training and defining what success means numerically.
- Training and comparison. Running the fine-tune and measuring against the untuned model on the held back set.
- Deployment and versioning. Serving the tuned model through your own API with the ability to fall back.
Fine-tuning teaches behaviour, retrieval supplies facts. Confusing those two is the single most common and most expensive mistake in this area.
How the process works, step by step
Fine-tuning usually takes four to eight weeks, most of it dataset work, and it begins with an assessment that frequently recommends against it.
-
Step 1
Free suitability review
We establish whether you need behaviour or knowledge, because only one of those is fine-tuning.
-
Step 2
Prompting and retrieval first
We try the cheaper approaches properly, and often the project ends here with the problem solved.
-
Step 3
Dataset built
Examples are gathered, cleaned and balanced, which is the largest part of the work by a wide margin.
-
Step 4
Evaluation held back
Test examples are separated before training, since measuring on seen data proves nothing.
-
Step 5
Trained and compared
The tuned model is measured against the untuned baseline and we report the difference honestly.
Quality assurance and how we work
The commitments attached to a technique that is usually the wrong one.
- We try prompting and retrieval first. They solve most cases at a fraction of the cost, and a fine-tune that could have been a prompt is money spent on maintenance you did not need.
- The evaluation set is held back before training. Measuring on data the model has already seen proves nothing, and it is the easiest way to report a false success.
- We compare against the untuned model. A fine-tune performing no better than a good prompt is not a success, and we will say so.
- We report when your dataset is not good enough. A small inconsistent set produces a worse model, and we say so rather than training on it and blaming the technique.
- We state the maintenance commitment upfront. Base models change and your fine-tune does not follow. Retraining is a recurring cost, not a one off project.
Frequently asked questions about AI model Fine-Tuning
Will fine-tuning make the model know our business?
No, and this is the most common misunderstanding. It teaches behaviour, not facts. For knowledge you want retrieval.
What is the difference exactly?
Fine-tuning changes how the model responds. Retrieval changes what it has in front of it. Confusing the two is the expensive mistake in this area.
Do you turn fine-tuning work down?
Frequently, because the client wants the model to know something rather than behave differently, and retrieval is cheaper and updateable.
When is fine-tuning actually right?
Strict output formats, specialist tone, or repetitive classification at volume where prompting has been tried and is not consistent enough.
How many examples do we need?
Usually hundreds at minimum, and quality matters more than quantity. A small clean set beats a large inconsistent one.
What if our examples are inconsistent?
We say so rather than training on them. An inconsistent dataset produces a worse model and then the technique gets blamed.
How do you measure success?
Against the untuned model on examples held back before training. Measuring on data the model has seen proves nothing.
What if it performs no better?
We report that. A fine-tune matching a good prompt is not a success and you should not pay to maintain it.
Does it need retraining later?
Yes. Base models change and your fine-tune does not follow automatically. That is a recurring cost we state before starting.
How long does a fine-tune take?
Four to eight weeks, most of it dataset work rather than training.
Why is the dataset the long part?
Because gathering, cleaning and balancing examples is genuinely the work. The training run itself is short.
Can it reduce our costs?
Sometimes considerably, since behaviour learned by the model does not need repeating in every prompt.
Does it lock us to one provider?
Largely, yes. A fine-tune is tied to its base model, which is a real trade off worth weighing before starting.
Can you fine-tune open models?
Yes, and self hosting is sometimes the better answer at volume. We compare honestly rather than assuming hosted.
Will it stop the model making things up?
No. Fine-tuning does not fix that, and expecting it to is another version of the behaviour and knowledge confusion.
Can it work on Swahili content?
Yes, and it is one of the cases where fine-tuning can genuinely help, though it depends on the quality of your examples.
Who owns the tuned model?
You do, along with the dataset, which is the more valuable asset of the two.
Can we use customer data as training examples?
Only with a lawful basis and appropriate handling, and we would raise that before it becomes a problem.
Is our training data kept private?
Yes. It stays under your control and is not reused for any other client or purpose.
What does fine-tuning cost?
Dataset size and training runs determine it, quoted after the suitability review, which may recommend against it.
Glossary of terms
Fine-tuning terms in plain language.
- Fine-tuning
- Adjusting a model on your examples so it adopts a style, format or behaviour.
- Retrieval
- Supplying facts to a model at request time, which is the correct approach for knowledge.
- Training set
- The examples the model learns from.
- Evaluation set
- Examples held back before training, used to measure honestly.
- Baseline
- The performance of the untuned model, without which improvement cannot be claimed.
- Overfitting
- A model that performs well on its training examples and poorly on anything else.
Related services in AI Development
Talk to us about AI model Fine-Tuning
Tell us what you want the model to do differently. If the answer is know our products, the technique you need is retrieval and it is cheaper.
Start your project