SheepDoctor: A knowledge graph enhanced large language model for sheep disease diagnosis

Traditional methods for diagnosing sheep diseases often fail to provide timely and accurate results, particularly in emergency situations, and public awareness of these diseases remains limited. This hinders effective prevention and control efforts. To address these challenges, a domain-specific large language model, SheepDoctor, was developed for sheep disease diagnosis. A comprehensive question-and-answer (Q&A) dataset was constructed using prompt techniques, resulting in 5,987 samples covering 207 sheep diseases. This dataset included detailed symptom descriptions, treatments, and related information, which were also structured into a knowledge graph. The pre-trained LLaMA2-13B-Chinese model was fine-tuned using Low-Rank Adaptation (LoRA), with integration of the knowledge graph to enhance its diagnostic capabilities and response accuracy. Evaluation using BLEU, ROUGE, and BERTScore metrics demonstrates that SheepDoctor outperforms general-purpose models such as GPT-4o and Kimi on sheep-related diagnostic tasks. The proposed method exhibits strong domain expertise and holds significant potential for improving the efficiency and reliability of sheep disease diagnosis.