Explained · Models
Fine-tuning
Fine-tuning takes a trained model and gives it a short second education on your own examples — your tone, your formats, your field — so the behaviour sticks without teaching anything from scratch. Think of hiring a brilliant generalist, then spending a week training them on your house style.
Where it breaksIt teaches behaviour, not facts, and people reach for it to fix the wrong problem. A model fine-tuned on your documents will start sounding like them without reliably knowing them, so answers get more confident and no more correct. And the moment your examples change or a better base model ships, you own a training pipeline you have to run again.
10,000× fewer trainable parameters reduction LoRA reports when adapting GPT-3 175B instead of retraining every weightHu et al., "LoRA: Low-Rank Adaptation of Large Language Models" · 2021-06-17
What changes inside the model
A model is billions of numbers, and those numbers were set by showing it an enormous amount of text and nudging them, over and over, towards predicting what came next. Fine-tuning is the same process, run again, on a very small pile of your own examples. You show it a few hundred or few thousand pairs — this input, that ideal output — and each one nudges the numbers a little further towards behaving that way. Nothing is added to the model; the weights it already had are simply moved. That distinction explains most of what fine-tuning can and cannot do. Moving weights is very good at shifting how the model responds, and unreliable at inserting facts, because a fact seen twenty times in fine-tuning is competing with everything the model absorbed during its original training.
Why it got cheap
Retraining every weight in a large model needs the same class of hardware the lab used, which put fine-tuning out of reach for almost everyone. The trick that changed this is to freeze the original weights entirely and train a small extra layer alongside them — a low-rank adapter, in the jargon — which learns the difference between the model's default behaviour and yours. Because that adapter is tiny relative to the model, the training run fits on ordinary hardware and finishes in minutes to hours rather than weeks. It also makes adapters swappable: one base model in memory, several personalities bolted on and off it, which is how providers offer per-customer fine-tunes without holding a separate copy of the model for every customer.
Where it breaks
Three ways, in rising order of expense. First, it is the wrong tool for knowledge — teams fine-tune on their documentation, get a model that adopts the house voice while still inventing the details, and conclude the technique failed when they simply needed retrieval. Second, the model can drift: train hard on a narrow set of examples and general ability degrades, so it gets better at your format and worse at reasoning about anything outside it. Third, and most often fatal in practice, you now own a pipeline. Your examples change, the base model is deprecated, a cheaper model ships that beats yours out of the box — and each of those means running the training, the evaluation and the deployment again. A prompt is edited in a minute; a fine-tune is a small standing commitment.
When it is actually worth it
Try the cheap things first, in order: a better prompt, a few examples in the prompt, then retrieval. Fine-tune when you have a behaviour you can demonstrate but not describe — a tone, a rigid output format, a domain style — and enough real examples to demonstrate it consistently. The other honest reason is unit economics: a fine-tuned small model that matches a large one on your narrow task can cut the per-request bill by an order of magnitude, and at high volume that pays for the pipeline. Both reasons need evals to confirm, because the failure mode of fine-tuning is a model that feels better and measures the same.