Fine-tuning

Fine-tuning is the process of taking a pretrained AI model — typically a foundation model like GPT-4, Llama 3, or Claude — and continuing its training on a smaller, task-specific dataset. The pretrained model already understands language, code, or images broadly; fine-tuning adapts it to a narrower domain (legal contracts, medical records, your support tickets) where general models struggle with vocabulary, conventions, or required output formats.

In machine learning terms, fine-tuning updates the model's weights using gradient descent on the new data, usually with a much lower learning rate than the original pretraining so the model doesn't "forget" its general knowledge. Common variants include supervised fine-tuning (SFT) on labeled examples, parameter-efficient methods like LoRA that only update a small subset of weights, and reinforcement learning from human feedback (RLHF) for aligning model behavior with preferences.

Fine-tuning is not always the right answer. For most business use cases, retrieval-augmented generation (RAG) — letting the model look things up at inference time — is cheaper and easier to update than fine-tuning. Fine-tuning wins when you need consistent output format, specialized terminology the base model doesn't know, or significantly lower per-token cost at scale.