Fine-tuning

Fine-tuning takes a pre-trained model and continues training it on your own smaller, task-specific dataset, adjusting the weights so the model internalizes your desired behavior. Instead of teaching new world knowledge, it mostly teaches form and skill: your output format, your tone, your domain’s jargon, your classification labels. Parameter-efficient methods like LoRA make this affordable by training small adapter layers instead of all the weights, and hosted services let you fine-tune API models by just uploading example conversations.

The reason this matters is the decision it forces. Most teams reach for fine-tuning too early when prompt engineering or RAG (retrieval-augmented generation) would solve the problem faster and cheaper. The working heuristic: prompt for behavior you can describe, retrieve for knowledge that changes, fine-tune for behavior you can only demonstrate — or when you need a small, cheap model to match a big one on a narrow task. Fine-tuning shines for consistent structured output, style matching, and cutting per-request costs by shrinking prompts or model size; it’s the wrong tool for keeping a model up to date on fresh facts.

In practice, the workflow is: collect a few hundred to a few thousand high-quality prompt/response pairs (usually as JSONL), run a training job — via a provider’s fine-tuning API or open-source tooling on Llama-class weights — then evaluate the resulting model against your baseline before switching production traffic to it.