Fine-tuning
Adjusting a pretrained model's parameters on additional data to specialize it for a particular task or style.
Fine-tuning takes an existing pretrained model and continues its training on a smaller, curated dataset — typically a few thousand to a few hundred thousand examples. The result is a specialized version of the base model that performs better on the specific task the fine-tuning data represented (a particular writing style, a domain vocabulary, a formatting convention).
The tradeoff: fine-tuning costs money and time, requires curated training data, produces a model version you have to maintain, and locks you into that version until you fine-tune again. For most use cases where people reach for fine-tuning, the right answer is either better prompting (free, iterative) or retrieval-augmented generation (grounds in current documents without a new model).
Fine-tuning is genuinely the right tool when: you have thousands of high-quality task-specific examples, the desired behavior can't be achieved by prompting, the base model consistently fails on your specific pattern, and the marginal quality gain justifies the ongoing cost. Most consumer use cases don't meet those criteria.