Learning rate
Beginner
How big a step the model takes each time it fixes a mistake. Too big and it wobbles; too small and it barely learns.
Novice
The size of each weight update during training. It’s the most important setting to get right: too high makes training unstable, too low wastes the run.
Expert
The step size multiplying the optimizer’s update, usually with warmup and a decay schedule. LoRA’s best rates sit about 10× above full fine-tuning’s (around 1e-4 vs 1e-5); sweep it rather than guess.
See also: Epoch, Overfitting.