What is fine-tuning in AI?

The plain idea

A pretrained model has already learned general patterns from a broad training run. Fine-tuning takes that model and continues training it on a smaller, curated set of examples specific to your task — customer-support replies in your brand’s voice, product descriptions in a fixed format, classification labels for your domain. The model’s internal parameters shift, in small ways, toward the patterns in your examples, so its future behaviour tends to follow them.

The result is a model that behaves differently by default, without you writing a longer prompt every time to coax it there. It is a change to the model itself, not to what the model is told at the moment of answering.

What fine-tuning is good for — and what it is not

Fine-tuning is genuinely effective for consistent style, format and task behaviour: getting a model to reliably answer in a house tone, to output a fixed structure, to use domain-specific terminology correctly, or to perform a narrow classification task well. If the problem is “the model can do this, but not consistently the way we need”, fine-tuning is a reasonable tool.

Fine-tuning is the wrong tool for injecting current or private facts into a model. It does not give the model a searchable memory of your documents, and it does not make yesterday’s policy update appear in tomorrow’s answers — the model’s knowledge is frozen at whatever was in the training examples, at the point training happened. That job belongs to retrieval, not training.

Fine-tuning versus RAG

This is the comparison worth getting right, and our explainer on retrieval-augmented generation covers it from the retrieval side. Fine-tuning changes the model’s behaviour and is frozen at training time — updating its knowledge means curating new examples and training again. Retrieval-augmented generation supplies fresh knowledge at query time by searching a document store and handing the model relevant passages to answer from, so it updates the moment a document does, and it can cite the sources it used.

A fair rule of thumb: if the problem is “the model needs to answer from changing, checkable facts”, start with retrieval. If the problem is “the model needs to consistently behave a certain way regardless of what it is asked”, consider fine-tuning. The two are not strict rivals — mature systems sometimes combine both, a fine-tuned model for behaviour with retrieval layered on top for facts — but reaching for fine-tuning to solve a knowledge problem is a common and avoidable mistake.

The honest costs and trade-offs

Fine-tuning is not free or simple. It needs curated training data — enough well-chosen examples to teach the behaviour without teaching noise — plus compute to run the training job and a genuine evaluation step to check the result actually improved rather than degraded on cases outside the training examples. There is a real risk of overfitting: a model that performs beautifully on examples like the ones it was trained on and worse on everything else.

And every future update to the desired behaviour means curating more examples and training again — there is no equivalent of simply editing a document. For many problems, a better-written prompt (see our explainer on prompt engineering) or a retrieval layer solves the underlying issue more cheaply and reversibly, and is worth trying first.

How cloud platforms offer it

Major cloud AI platforms offer managed fine-tuning so teams do not need to run their own training infrastructure — you supply curated examples, the platform handles the training job, and the resulting customised model becomes available to call like any other. Amazon Bedrock is the AWS expression of this pattern, offering model customisation as a managed capability alongside its other generative-AI services, without requiring teams to manage the underlying training infrastructure themselves.

Fine-tuning in certification study

Fine-tuning falls under model customisation on the AWS Certified AI Practitioner (AIF-C01), typically presented alongside prompt engineering and retrieval-augmented generation as the three main ways to adapt a foundation model’s behaviour. The exam expects a conceptual grasp of when fine-tuning is the appropriate choice versus prompting or retrieval, in a described scenario — precisely the comparison covered above — rather than the training mechanics themselves.