Fine-tuning is a technique for adapting a pre-trained foundation model to a specific domain, task, or style by continuing to train it on a curated dataset relevant to that use case. The model starts with everything it learned during its original training and then adjusts its weights based on the new examples, learning to produce outputs that better match the target distribution.
Common reasons to fine-tune include improving a model's performance on a specialized domain where general training data was sparse, teaching a model a specific output format or persona, reducing the need for lengthy system prompts by baking instructions into the model's weights, and improving consistency in structured tasks like data extraction or classification.
Fine-tuning is often considered before it should be. In most enterprise use cases, the combination of well-crafted system prompts, retrieval-augmented generation, and few-shot examples in the context achieves performance close to what fine-tuning would produce, at a fraction of the cost and maintenance burden. Fine-tuning should be pursued when there is evidence that simpler approaches have a ceiling that the use case cannot tolerate.
When fine-tuning is the right choice, the quality of the training data matters more than its quantity. A few hundred high-quality, carefully curated examples consistently outperform thousands of noisy or inconsistent ones. The fine-tuned model also needs to be re-evaluated and potentially re-trained as the base model is updated by its provider, which adds an ongoing operational responsibility.