Full tuning moves every weight, which costs memory and produces a whole model per job. LoRA trains small add-on matrices instead, so a job is cheap and several jobs sit on one base model. The trade is capacity: a narrow adapter learns a narrow job well and a broad one poorly.
About 1 min of reading
Check the plateau.
The curve
Before you run it, say out loud what it will print. Then open the answer.
Read the first example one line at a time. A line that prints shows what appears on screen; the rest set things up quietly.
A format or style job fits a small adapter. A job that needs the model to learn a whole new domain is closer to full tuning, and forcing it into a tiny adapter produces a quiet partial failure.
Merging an adapter into the base makes one model and loses the ability to switch. Keep adapters separate until a serving constraint forces the merge.
One base model with several adapters is several jobs. Switching the adapter switches the job, which is what makes a set of small tuned behaviours practical where full models would be many gigabytes each.
The adapter's rank is how much it can learn. A job that needs a large change in behaviour with a tiny rank learns a weak version of it, and the symptom is a held out score that plateaus early.
Which layers get the adapter matters. A job that changes output shape usually needs the projection layers; a job that changes knowledge needs more, and more is closer to full tuning.
Each one is defined here on this site. Open the entry for the full definition and the other words it points at.