Overfitting

Pick the sensible `num_train_epochs` for a typical instruction-SFT run on a curated dataset

Name the signal early stopping watches, and explain the 'patience' knob

Train loss drops while eval loss climbs at step 1200: name it and the standard fixes

Misconception: 'higher LoRA rank always wins on benchmarks'

When does targeting all-linear with LoRA pay off vs. just W_q+W_v?

Pick a LoRA rank for a domain task and justify why r=512 isn't a free upgrade

Match LoRA rank to task type