Small language models learn enhanced reasoning skills from medical textbooks

Hyunjae Kim, Hyeon Hwang, Jiwoo Lee, Sihyeon Park, Dain Kim, Taewhoo Lee, Chanwoong Yoon, Jiwoong Sohn, Jungwoo Park, Olga Reykhart, Thomas Fetherston, Donghee Choi, Soo Heon Kwak, Qingyu Chen, Jaewoo Kang

Abstract: Small language models (SLM) offer promise for medical applications by addressing the privacy and hardware constraints of large language models; however, their limited parameters (often fewer than ten billion) hinder multi-step reasoning for complex medical tasks. This study presents Meerkat, a new family of medical SLMs designed to be lightweight while enhancing reasoning capabilities. We begin by designing an effective and efficient training method. This involves extracting high-quality chain-of-thought reasoning paths from 18 medical textbooks, which are then combined with diverse instruction-following datasets within the medical domain, totaling 441K training examples. Fine-tuning was conducted on open-source SLMs using this curated dataset. Our Meerkat-7B and Meerkat-8B models outperformed their counterparts by 22.3% and 10.6% across six exam datasets, respectively. They also improved scores on the NEJM Case Challenge from 7 to 16 and from 13 to 20, surpassing the human score of 13.7. Additionally, they demonstrated superiority in expert evaluations, excelling in all metrics—completeness, factuality, clarity, and logical consistency—of reasoning abilities.