Innovative Design of Large Language Model in the Medical Field Based on chip-PromptCBLUE

This article introduces the research content and results based on the CHIP-PromptCBLUE (Chinese Biomedical Language Understanding Evaluation) benchmark task. PromptCBLUE promotes research on large language models for medicine. The benchmark can evaluate Chinese language models’ multi-tasking abilities across various medical tasks, including 18 task types such as medical entity recognition, medical text classification, medical language inference, and medical content generation. It requires completing all tasks using just one large language model, necessitating efficient fine-tuning methods and keeping parameters within 1% of the model size. To address this, we propose a method. First, we greatly improved model performance through data augmentation. We then further amplified model capabilities using an innovative entity loss optimization of the large model’s loss function. Using this method, we achieved a score of 71.3822 in the chip-PromptCBLUE general track. This research provides new ideas for advancing large language models in the medical field.