“For short training runs of approximately 100 steps or less, the optimal learning rate multiplier for Low-rank adaptation (LoRA) over full fine-tuning is around 15x.”