Fine-tuning is teaching a model new tricks

Base models like Llama 3 or Mistral are general — they know a little about everything, like a college grad who's read a lot but hasn't picked a trade. Fine-tuning is the next step: you take one of those models and train it on your own data so it learns a specific way of speaking, a specific domain, a specific style.

The way it works is simple. You take a base model, you give it a bunch of examples — question and answer pairs, or paragraphs — and the model tweaks its internal weights to match the pattern. It's not a new model from scratch. It's the same architecture, just nudged.

Think of it like this: you already know how to speak English. Now your abuela wants you to explain the doctor's visit to tía Rosa in the way she'd actually talk — no jargon, the right words, the way we do. You don't relearn language. You shift the register. Fine-tuning does that shift for a model.

Fine-tuning is different from prompting or RAG. Prompting is just telling the model what you want in the moment. RAG is pulling in outside documents at inference time. Fine-tuning changes the model itself — it permanently absorbs the pattern.

It's popular because it's cheap compared to building a new model and fast enough to do in a day with the right tools. But it has limits. You can't fine-tune a model to know facts it never saw in the training data — for that you still need retrieval. And over-fine-tuning can make the model forget how to reason on general tasks.

Why this matters for us:
Use fine-tuning for style and domain, not for fresh facts — pair it with retrieval when you need accuracy.