Rethinking Data Selection for Supervised Fine-Tuning

Although supervised finetuning (SFT) has emerged as an essential technique to align large language models with humans, it is considered superficial, with style learning being its nature. At the same time, recent works indicate the importance of data selection for SFT, showing that finetuning with high-quality and diverse subsets of the original dataset leads to superior downstream performance. In this work, we rethink the intuition behind data selection for SFT. Considering SFT is superficial, we propose that essential demonstrations for …

cs.cl data dataset diverse fine-tuning finetuning humans importance language language models large language large language models leads nature quality sft style supervised fine-tuning