“Low-rank adaptation (LoRA) is less tolerant of large batch sizes than full fine-tuning, paying a larger penalty in loss as batch size increases.”