Fine-tuning

Continuing to train a pre-trained model on a smaller, task-specific dataset so it specializes in a particular domain or behavior.

Fine-tuning is the process of taking an already pre-trained model — like Llama, GPT, or a base BERT — and continuing to train it on a much smaller dataset that reflects a specific task, domain, or style. The model keeps its general language abilities but adjusts its weights to specialize. It matters because training a foundation model from scratch costs millions of dollars, while fine-tuning can be done on a few thousand examples in hours. This is how teams turn a general-purpose LLM into a customer-support bot in their company's voice, a medical-summary assistant, or a code generator for an internal language. Fine-tuning is also how RLHF (reinforcement learning from human feedback) shapes raw base models into helpful chat assistants like ChatGPT or Claude. The analogy: a pre-trained model is like a university graduate who knows a bit of everything. Fine-tuning is the on-the-job training that turns them into a specialist — a paralegal, a radiologist, a Python reviewer. They didn't relearn English; they just learned the patterns of their new role. In practice, full fine-tuning updates every weight in the model, which is expensive. Most teams now use parameter-efficient methods like LoRA or QLoRA that freeze the base model and train only small adapter layers — cheaper, faster, and easy to swap. For many use cases, prompt engineering or RAG (retrieval-augmented generation) handles the job without any training at all, so fine-tuning is usually a last resort when you need consistent style, structured output, or domain knowledge that prompts can't reliably elicit. Related concepts to explore next: LoRA, RLHF, instruction tuning, transfer learning, RAG, pre-training.