Fine-tuning

Fine-tuning is further training of an already pre-trained model on a smaller, task-specific dataset so it adapts to a domain, format, or style. The base model supplies general language ability, and fine-tuning adjusts weights to shift behaviour, which is why a few thousand good examples can outperform a very long prompt. Full fine-tuning updates every parameter and is expensive, so most work uses parameter-efficient methods. LoRA trains small low-rank adapter matrices and leaves the base weights frozen, and QLoRA adds quantization so a large model can be adapted on a single GPU; adapters are small files that can be swapped per customer. Typical uses are enforcing a strict output schema, matching a house tone, handling specialist vocabulary, and distilling a large model's behaviour into a cheaper small one. Fine-tuning teaches behaviour, not facts, and this is the usual mistake. For information that changes, retrieval-augmented generation is the right tool, since a fine-tuned model still needs retraining when the underlying data moves. Poorly curated data also causes catastrophic forgetting, where the model improves on the narrow task and gets worse at everything else.