Fine-tuning

Fine-tuning takes a model that has already been trained on a broad, general dataset and continues training it on a smaller, more specific set of examples, adjusting its weights so it performs better on a particular task, tone, or domain than the general model would on its own.

It differs from prompt engineering, which shapes behavior only through the instructions given at inference time without touching the model's weights. Fine-tuning is more durable and can encode patterns too complex or too numerous to fit in a prompt, but it requires curated training data, compute, and careful hyperparameter tuning to avoid overfitting to the fine-tuning set at the expense of general capability, an effect sometimes called catastrophic forgetting.

Fine-tuning matters when prompting alone cannot reliably produce the needed behavior, style, or domain accuracy, though many teams now try retrieval-augmented generation first since it avoids retraining. Common pitfalls include fine-tuning on too little or unrepresentative data, skipping held-out evaluation so regressions go unnoticed, and treating fine-tuning as a way to add new factual knowledge, when it is generally better suited to adjusting style and behavior. Operationally it needs the same MLOps discipline as any model retraining.