Advanced PromptCBLUE Performance: A Novel Approach Leveraging Large Language Models

Hong‐Qing Ling,Congcong Ge,Jie Wang,Fuliang Quan,Jianping Zeng
2023-01-01
Abstract:PromptCBLUE is a Chinese medical NLP benchmark that converts sixteen different medical NLP tasks into text generation task through prompt learning. It requires completing all tasks with just one large language model backbone and keeping the fine-tuning parameters under 1% of the model size, posing huge challenges. To address this issue, we have proposed an innovative approach. Initially a two-stage fine-tuning process is employed. For intricate tasks, a task-decomposition prompting construction method is introduced, enhancing algorithmic efficacy. Moreover, the integration of focus loss is explored to further amplify algorithm performance. Using this approach, a score of 72.28 was achieved in the CCKS2023 PromptCBLUE general competition track. This research contributes new insights to the advancement of large language models in the medical field.
What problem does this paper attempt to address?