“The optimal learning rate for Low-rank adaptation (LoRA) is consistently 10 times higher than for full fine-tuning (FullFT) across both supervised learning and reinforcement learning.”