The chef analogy
Imagine a chef trained classically in France. They can cook almost anything competently — that is the base model, a generalist with broad skill. Now send that chef to Japan for an intensive course in sushi. They do not forget how to make a sauce or run a kitchen; they simply gain a sharp new specialty on top of everything they already knew. Fine-tuning is exactly that: you take a capable general model and train it a little further on examples from a narrow domain, so it becomes noticeably better at one kind of task without losing its broad foundation.
What fine-tuning actually does
A base model is a giant set of numbers — its weights — learned from general text. Fine-tuning continues that training on a focused dataset of your own examples, nudging the weights so the model’s default behaviour shifts toward what you showed it. The key difference from prompting is permanence. A prompt influences a single conversation and must be repeated every time. Fine-tuning bakes the behaviour into the model, so it applies automatically on every call — shorter prompts, more consistent output, no re-explaining your house style each time.
When to reach for it
Fine-tuning shines when you need consistency at scale. If you want every response in a specific tone, every output in a strict format (always valid JSON, always a particular structure), or a reliable style across thousands of requests, fine-tuning enforces it far better than hoping a prompt holds. It is also worth it when your prompts have ballooned with examples and instructions — folding those into the model can cut cost and latency. For one-off tasks, fast-changing requirements, or small volumes, a well-written prompt is cheaper and quicker.
What it is not good for
A common trap is treating fine-tuning as a way to teach the model new facts. It is weak and risky for that. Facts learned this way are hard to update, easy to get wrong, and prone to being recalled inaccurately. When you need current, citable knowledge — product details, recent events, internal documents — use retrieval-augmented generation (RAG) instead: fetch the relevant documents and hand them to the model at query time. Think of the split as fine-tuning teaches behaviour, RAG provides knowledge.
Doing it without breaking the chef
The danger of fine-tuning is overfitting — drilling the model so hard on narrow examples that it loses its general competence, like a chef who forgets everything but sushi. The defences are straightforward: use high-quality, varied examples, keep the dataset clean, train gently, and always hold out a test set to check the model still performs well beyond your narrow data. Modern efficient methods such as LoRA make this cheaper and safer by adjusting only a small slice of the model. Done with care, fine-tuning gives you a specialist that still remembers everything it knew before.