Synthesizing Instruction-Tuning Datasets with Contrastive Decoding
arXiv:2604.13538v1 Announce Type: new
Abstract: Using responses generated by high-performing large language models (LLMs) for instruction tuning has become a widely adopted approach. However, the existing literature overlooks a property of LLM-generat…