Fine-tuning

A practical guide to understanding what fine-tuning is and when it makes sense for AI product development.

What fine-tuning does to an AI model, when it is worth doing, and what product and design teams need to know before commissioning it.

22 May 20265 min read

What it is

glossaryFine-tuningFine-tuning is the process of further training a pre-trained model on specific data to improve performance for a particular task.Open glossary term is the glossaryProcessA process is a defined sequence of steps used to achieve a specific outcome.Open glossary term of taking a pre-trained glossaryModelA model is a system or representation used to process data and generate outputs, often trained to perform specific tasks.Open glossary term and training it further on a smaller, specific dataset to improve its performance on a particular task or domain.

Rather than training a glossaryModelA model is a system or representation used to process data and generate outputs, often trained to perform specific tasks.Open glossary term from scratch — which requires enormous resources — glossaryFine-tuningFine-tuning is the process of further training a pre-trained model on specific data to improve performance for a particular task.Open glossary term adjusts the existing model's glossaryBehaviourBehaviour refers to how users interact with a system, including actions, patterns, and responses.Open glossary term to better suit a specific context.

A general-purpose glossaryModelA model is a system or representation used to process data and generate outputs, often trained to perform specific tasks.Open glossary term might be fine-tuned to write in a particular brand's tone, handle specific customer support queries, or generate content in a specialist domain like legal or medical writing.

The result is a glossaryModelA model is a system or representation used to process data and generate outputs, often trained to perform specific tasks.Open glossary term that performs better on the target task than a general model prompted in the usual way.

glossaryFine-tuningFine-tuning is the process of further training a pre-trained model on specific data to improve performance for a particular task.Open glossary term is not always necessary. Many tasks can be handled effectively through well-designed glossaryPromptA prompt is the input or instruction given to an AI system to guide its output or response.Open glossary term without touching the glossaryModelA model is a system or representation used to process data and generate outputs, often trained to perform specific tasks.Open glossary term itself.

When to use it

Understand when glossaryFine-tuningFine-tuning is the process of further training a pre-trained model on specific data to improve performance for a particular task.Open glossary term is worth the investment.

It is most relevant when:

A general model consistently underperforms on a specific task despite good prompting

You have a large volume of high-quality, labelled training examples

Consistent tone, style, or domain accuracy is critical

Latency or cost savings from shorter prompts would be valuable at scale

The task is well-defined and unlikely to change frequently

It is less relevant when:

Good prompt engineering can achieve the required results

Training data is limited, low quality, or difficult to label

The task or requirements change frequently

Budget and time for training and evaluation are constrained

Key takeaway

Fine-tuning is a serious investment. Before commissioning it, be confident that prompt engineering alone cannot solve the problem.

How it works

Understand the basic mechanism. glossaryFine-tuningFine-tuning is the process of further training a pre-trained model on specific data to improve performance for a particular task.Open glossary term works by continuing the training glossaryProcessA process is a defined sequence of steps used to achieve a specific outcome.Open glossary term on a curated glossaryDatasetA dataset is a structured collection of data used for analysis, training models, or processing.Open glossary term relevant to the target task.

The glossaryModelA model is a system or representation used to process data and generate outputs, often trained to perform specific tasks.Open glossary term is exposed to many examples of the input-output pairs you want it to learn — for example, customer queries paired with ideal glossaryResponseA response is the data or result returned by a server after receiving a request.Open glossary term, or product descriptions paired with desired rewrites.

Through this glossaryProcessA process is a defined sequence of steps used to achieve a specific outcome.Open glossary term, the glossaryModelA model is a system or representation used to process data and generate outputs, often trained to perform specific tasks.Open glossary term's parameters are adjusted to make it more likely to produce outputs that match the examples it has been shown.

The fine-tuned glossaryModelA model is a system or representation used to process data and generate outputs, often trained to perform specific tasks.Open glossary term retains its general glossaryCapabilityCapability refers to an organisation’s ability to perform a specific function or deliver a particular outcome.Open glossary term from the original training while becoming more reliable on the specific tasks it has been fine-tuned for.

What this means for designers and product teams. glossaryFine-tuningFine-tuning is the process of further training a pre-trained model on specific data to improve performance for a particular task.Open glossary term decisions start with glossaryDataData is raw information collected and stored for analysis, processing, or decision-making.Open glossary term. The quality, quantity, and glossaryConsistencyConsistency is the use of uniform patterns, behaviours, and visual elements across a product to create familiarity and predictability. It helps users learn once and apply that knowledge throughout the experience.Open glossary term of the training examples directly determine the quality of the fine-tuned model.

As a designer or product person, your most important contribution is defining what good output looks like and helping to curate or evaluate the guideTraining DataWhat training data is, how it shapes what an AI model knows and assumes, and what product and design teams need to understand about its role in AI product quality.Open guide.

Fine-tuned glossaryModelA model is a system or representation used to process data and generate outputs, often trained to perform specific tasks.Open glossary term also need to be evaluated properly — just because a model has been fine-tuned does not mean it is reliable. Testing on unseen examples is essential.

What to look for

Focus on:

Training data quality — whether examples are accurate, consistent, and representative

Task definition clarity — whether the target behaviour is clearly specified

Evaluation rigour — whether the model is tested against real-world inputs after fine-tuning

Overfitting risk — whether the model has become too narrowly optimised and fails on variation

Ongoing maintenance — whether the training data and model will be kept current

Where it goes wrong

Most issues come from: A fine-tuned glossaryModelA model is a system or representation used to process data and generate outputs, often trained to perform specific tasks.Open glossary term trained on poor glossaryDataData is raw information collected and stored for analysis, processing, or decision-making.Open glossary term will consistently produce poor outputs — at scale.

Insufficient or low-quality training data

Poorly defined success criteria for what good output looks like

No evaluation process after fine-tuning is complete

Overfitting to the training examples and failing on real variation

Treating fine-tuning as a one-time task rather than an ongoing process

What you get from it

Understanding glossaryFine-tuningFine-tuning is the process of further training a pre-trained model on specific data to improve performance for a particular task.Open glossary term gives you:

A clearer basis for deciding when fine-tuning is and is not the right approach

Better ability to contribute to the data curation and evaluation process

More realistic expectations about what fine-tuning can and cannot achieve

A stronger brief for working with engineers and ML teams

Key takeaway

Fine-tuning is powerful when the task is clear and the data is strong. It is expensive and slow when either of those conditions is not met.

FAQ

Common questions

A few practical answers to the questions that usually come up around this method.

What is fine-tuning in simple terms?

It is the glossaryProcessA process is a defined sequence of steps used to achieve a specific outcome.Open glossary term of taking an existing glossaryModelA model is a system or representation used to process data and generate outputs, often trained to perform specific tasks.Open glossary term and training it further on a specific set of examples to make it better at a particular task. Rather than building a new model from scratch, you are adjusting an existing one to suit your needs.

How is fine-tuning different from prompting?

Prompting guides an existing glossaryModelA model is a system or representation used to process data and generate outputs, often trained to perform specific tasks.Open glossary term through instructions at the point of use. glossaryFine-tuningFine-tuning is the process of further training a pre-trained model on specific data to improve performance for a particular task.Open glossary term changes the model itself by training it on new examples. Prompting is faster and cheaper to iterate — fine-tuning is more powerful but requires significant time, glossaryDataData is raw information collected and stored for analysis, processing, or decision-making.Open glossary term, and cost.

How much data do you need to fine-tune a model?

It depends on the task and the glossaryModelA model is a system or representation used to process data and generate outputs, often trained to perform specific tasks.Open glossary term, but quality matters more than quantity. A few hundred high-quality, well-labelled examples can be enough for some tasks. For complex or varied tasks, you may need thousands. Poorly labelled glossaryDataData is raw information collected and stored for analysis, processing, or decision-making.Open glossary term in large volumes will produce worse results than a smaller, well-curated set.

Is fine-tuning permanent?

A fine-tuned glossaryModelA model is a system or representation used to process data and generate outputs, often trained to perform specific tasks.Open glossary term reflects the glossaryDataData is raw information collected and stored for analysis, processing, or decision-making.Open glossary term it was trained on at a point in time. If the task, content, or requirements change, the model will need to be re-evaluated and potentially retrained. It is not a one-time fix.

Do designers need to be involved in fine-tuning?

Yes, in the sense that defining what good output looks like and evaluating whether the fine-tuned glossaryModelA model is a system or representation used to process data and generate outputs, often trained to perform specific tasks.Open glossary term meets user needs are design and product decisions. The technical training is for engineers and ML teams, but the quality criteria, the use cases, and the evaluation of real-world glossaryPerformancePerformance refers to how quickly and efficiently a system responds to user actions and processes tasks.Open glossary term are squarely in the product and design domain.

Quick take

If a general AI model is not performing well enough for your specific use case, fine-tuning adapts it to your needs — but it is not always the right answer.

Related Services

Artificial Intelligence