Search

What are you looking for?

Search our services, use cases and practical insights.

Enter at least 2 characters

Popular starting points

AI explained simply

What is fine-tuning – and when is it useful?

Fine-tuning continues training an existing model on selected examples to produce a desired behaviour. It is an optimisation method, not a simple store for current business knowledge.

The short answer

Fine-tuning is additional training of a pretrained model on a curated dataset so that it performs particular tasks, formats, terminology or behavioural patterns more consistently.

In brief

  • Fine-tuning changes model parameters; a prompt only changes the context of one request.
  • It is more suitable for repeatable behaviour than for frequently changing facts.
  • The quality and representativeness of training examples matter more than volume alone.
  • Separate evals are needed before and after training, together with a plan for versions and operation.

What happens during fine-tuning?

A base model is further optimised using examples of inputs and desired outputs. This increases the likelihood that it handles similar tasks in the demonstrated style or format. Training data is not stored like rows in a database and cannot be queried as an authoritative source.

  • Select a base model and supported training method
  • Clean examples and convert them to the expected format
  • A training run produces a new model version
  • Independent cases test benefit and unintended effects

When can fine-tuning be useful?

Fine-tuning may be worthwhile when a measurable task remains systematically inconsistent despite a good prompt and suitable context. It can be especially useful for stable classifications, specialised output formats or a large volume of similar requests.

  • Apply domain terminology or recurring categories more consistently
  • Produce a complex output format more reliably
  • Shorten long example instructions and reduce runtime cost
  • Optimise a smaller model for a narrowly bounded task

Which approach solves which problem?

Current product data, policies and customer details should come from controlled sources. The following decision aid separates four approaches that are often confused. They can be combined, but they change different parts of a solution.

Which approach solves which problem?
ApproachSuitable whenNot the primary solution forHow benefit is demonstrated
PromptingThe task, tone or output format must change quicklyNew facts, technical permissions or exact calculationsBetter results on the same test set without model training
RAGChanging knowledge with citations and user permissions is requiredLearning persistent behaviour or executing live transactionsThe correct authorised passage is retrieved and cited accurately
API or toolExact live values, calculations or controlled actions are neededOpen-ended interpretation of unstructured contentParameters, result and failure behaviour are technically verifiable
Fine-tuningStable behaviour remains inconsistent despite a strong promptFrequently changing knowledge, citations or permissionsA new model version measurably and consistently reduces defined errors

How do you make an evidence-based decision?

A training project needs a measurable baseline. The team compares the base model, improved prompting, RAG or tools and fine-tuning on the same test set. Only a demonstrated advantage justifies data preparation and additional model operations.

  1. Step 1

    Demonstrate failure patterns

    Show which recurring errors occur in the current solution.

  2. Step 2

    Test simpler options

    Try prompt, context, structured output and a suitable model first.

  3. Step 3

    Separate data

    Build non-overlapping training, validation and test sets.

  4. Step 4

    Assess total cost

    Include training, evaluation, versioning, privacy and operation.

Example from day-to-day business

Example: classifying technical support cases

A company has thousands of reviewed tickets with stable issue categories. Prompting produces too many inconsistent labels for rare domain terms. Fine-tuning on cleaned historical examples improves classification by a smaller model. Current product versions and resolution steps still come from RAG; the trained model does not replace those sources.

What to remember

Fine-tuning is targeted optimisation after a measurable need has been shown. Use it for stable behaviour, not as a shortcut for data integration, permissions or current knowledge.

Sources and further reading

These primary sources provide further detail on definitions, technical foundations or responsible use.

Content reviewed

Would you like to apply this to your situation?

Together, we clarify what makes sense for your process, data and systems – in plain language and without unnecessary complexity.

Discuss Your Project