Qichen Ye, Junling Liu, Dading Chong, Peilin Zhou, Yining Hua, Andrew Liu, Xuxin Cheng
Integrating large language models (LLMs) into healthcare presents potential but faces challenges. Directly pre-training LLMs for domains like medicine is resource-heavy and sometimes unfeasible. Sole reliance on Supervised Fine-tuning (SFT) can result in overconfident predictions and may not tap into domainspecific insights. Addressing these challenges, we present a multi-stage training method combining Domain-specific Continued Pre-training (DCPT), SFT, and Direct Preference Optimization (DPO). A notable contribution of our study is the introduction of a 3Gb Chinese Medicine (ChiMed) dataset, encompassing medical question answering, plain texts, knowledge graphs, and dialogues, segmented into three training stages. The medical LLM trained with our pipeline, Qilin-Med, exhibits significant performance boosts.
- training scripts
- training data
- models
Our ChiMed dataset is avalibale at here, which containing the following three parts:
- ChiMed-Pretrain(
cpt.txt), - ChiMed-SFT(
sft.jsonl) - ChiMed-DPO(
dpo.json)
Put the CHiMed-Pretrain data (i.e., cpt.txt) at data/pretrain/, then run the following scripts.
bash run_pt.sh
Put the CHiMed-SFT data (i.e., sft.jsonl) at data/sft/, then run the following scripts.
bash run_sft.sh
Put the CHiMed-DPO data (i.e., dpo.json) at data/dpo/, then using scripts/merge_peft_adapter.py to merge the sft adapter with Qilin-Med-Pretrained, then put the resulting model to checkpoints/Qilin-Med-SFT-merged. Finally run the following scripts.
bash run_dpo.sh
@misc{ye2023qilinmed,
title={Qilin-Med: Multi-stage Knowledge Injection Advanced Medical Large Language Model},
author={Qichen Ye and Junling Liu and Dading Chong and Peilin Zhou and Yining Hua and Andrew Liu},
year={2023},
eprint={2310.09089},
archivePrefix={arXiv},
primaryClass={cs.CL}
}
Many thanks to the following awesome works!