Fine-Tuning the Llama2 Large Language Model Using Books on the Diagnosis and Treatment of Musculoskeletal System in Physical Therapy

Background: In physical therapy, accurate information on musculoskeletal diagnosis and treatment is essential for patient care. Generative language models use machine learning to perform tasks like text generation and question answering, mimicking human language understanding. Purpose: This study aimed to fine-tune the Llama2 language model using text data from books on the diagnosis and treatment of the musculoskeletal system in physical therapy, and to compare it to the base model to evaluate its usability in medical fields. Study design: Technical evaluation study Methods: The Llama2-13B Chat model was fine-tuned using text from books on musculoskeletal diagnosis and treatment in physical therapy and trained on 1.2 million tokens. Responses were compared to the base model based on relevance and specialization to clinical questions. Results: Compared to the base model, the fine-tuned model consistently generated answers specific to the musculoskeletal system diagnosis and treatment, demonstrating improved understanding of the specialized domain. Conclusions: The model fine-tuned for musculoskeletal diagnosis and treatment books provided more detailed information related to musculoskeletal topics, and the use of this fine-tuned model could be helpful in medical education and the acquisition of specialized knowledge.