What is fine-tuning? Continue training a model so a behavior becomes part of the model itself

Suppose you hire someone who is already an excellent writer.

You could explain your house style from scratch every time you assign an article. Or you could train that person repeatedly on your examples until the style becomes habitual.

That second idea is a useful analogy for fine-tuning.

Fine-tuning continues training an existing model

A foundation model has already learned broad patterns from large-scale training.

Fine-tuning uses additional examples to update some or all model parameters so that particular behaviors become more likely.

Depending on the method, you may train the full model or use parameter-efficient techniques that update a much smaller set of learned values.

Prompting and fine-tuning change different things

A prompt changes the instructions and context for this request.

Fine-tuning changes the model’s learned parameters so the behavior can persist across requests without repeating as much instruction.

That makes fine-tuning useful for recurring patterns such as:

• a specific output structure,
• classification behavior,
• domain terminology,
• tool-call style,
• repeated tone or format.

Fine-tuning is not a database

If your product needs today’s exchange rate or a policy updated every week, embedding those facts into weights can be inconvenient and expensive.

RAG from Lesson 023 is often a better fit for frequently changing or auditable knowledge because the source documents remain external and replaceable.

A simple rule of thumb:

• change what the model knows at request time → consider RAG/context,
• change how the model repeatedly behaves → consider fine-tuning.

The boundary is not absolute, but it is a useful starting point.

Training data quality matters more than file count

A thousand inconsistent examples can teach conflicting behavior.

A smaller set of clear, representative examples may be more useful.

You should define:

• what the input looks like,
• what a correct output looks like,
• difficult edge cases,
• what should not change,
• a held-out evaluation set.

Do not evaluate only on the same examples used for training.

Fine-tuning can also teach mistakes

If labels are wrong, style is inconsistent or sensitive information leaks into training examples, the model can learn those problems.

Fine-tuning is not a “make model smarter” button. It optimizes toward the examples and objective you provide.

Cost and maintenance matter

Training has a cost, and a fine-tuned model can become stale as requirements change.

Before training, test whether a stronger prompt, structured output, tool use or RAG already solves the problem.

Fine-tuning is most valuable when the improvement repeats at scale.

One thing to remember

Fine-tuning updates a model through additional training so a desired behavior becomes part of its learned parameters, rather than only an instruction in one prompt.

Lesson 025 moves from model customization to model location: what changes when you run a local LLM on your own hardware?