Supervised fine-tuning (SFT) is widely adopted for tailoring large language models (LLMs)
to specific downstream tasks. However, the substantial computational demands of LLMs …