基于多轮对话增强的婴儿培养箱不良事件问答大模型微调OA

This paper addresses the challenges of knowledge domain overlap and the large volume of reports in monitoring ad-verse events related to infant incubators,which make manual processing time-consuming and labor-intensive.It proposes a fine-tuned large-scale question-answering model for infant incubator adverse events based on multi-round dialogue enhancement,aiming to assist monitoring personnel in quickly handling related adverse events.The study utilizes adverse event data provided by a specific province,combined with the characteristics of infant incubator adverse event reports.It innovatively constructs a fully annotated multi-round dialogue dataset conducive to training models for infant incubator adverse events.It also integrates hi-erarchical position encoding and dynamic attention masking to improve the sensitivity of each round and resistance to cross-round interference.Additionally,using ChatGLM3-6B as the base model,the pre-trained large language model is fine-tuned with LoRA technology to enable a deep understanding of the working principles of infant incubators and the features of adverse events.Experimental results show that the fine-tuned model significantly surpasses the baseline in Rouge Score and BLEU met-rics,with improvements of 15.9%points and 16.5%points,respectively.In the human evaluation phase,it also demonstrated superior performance in question-answering tasks,with a 14.3%increase in scores.This study effectively reduces the workload of monitoring personnel and offers new insights and methods for applying large language models in the field of medical device monitoring.