Lesson 2 of 6
How fine-tuning works (SFT, LoRA)
7 min read
You show a model a few hundred of your best answers. How does that actually change what it says next time?
Show, don't tell
The workhorse is supervised fine-tuning (SFT): you hand the base model many example pairs — a prompt and the ideal response — and training nudges its weights so it's more likely to produce responses like yours. Batch by batch, the share of outputs matching your target style climbs.
SFT doesn't add new facts — it reshapes tendencies. Every batch of examples tilts the model a little further toward the behavior you demonstrated.
LoRA: tune a small add-on, not the whole model
Retraining every weight is expensive and memory-hungry. LoRA (low-rank adaptation) freezes the original model and trains a tiny set of extra weights that sit alongside it — often under 1% of the size. You get most of the benefit of full fine-tuning for a fraction of the compute, and you can keep many small adapters for one base model.
LoRA changes the economics: the same behavior gain, a fraction of the compute and storage. It's why fine-tuning is now practical on modest hardware.
Full fine-tuning and LoRA aim at the same target — LoRA just gets there cheaper. For most teams LoRA is the default; full fine-tuning is reserved for the rare case that needs it.
The shape of it
- —SFT trains on prompt→response pairs, shifting tendencies toward your examples.
- —It reshapes behavior, not knowledge — no new facts are added.
- —LoRA trains a small adapter, not the whole model — most of the gain, far less cost.
What does LoRA change compared to full supervised fine-tuning?
Continue in the app
Take the whole Fine-tuning & Customization course — tracked
Get your personalized path, progress and streaks in the app — this lesson and every next one, in order.