IC-BERT: An Instruction Classifier Model Alleviates the Hallucination of Large Language Models in Traditional Chinese Medicine

Large language models (LLMs) may occasionally provide inaccurate answers to questions and answers in specialized domains. Therefore, we implemented a text filter based on Bidirectional Encoder Representations from Transformers (BERT) to screen questions, limiting the scope of the questions that the large model can answer. Although fine-tuning can equip large language models with some ability to recognize questions, confusion may still arise when facing other questions similar to those in the fine-tuning dataset, potentially leading to incorrect answers even for questions that might have been responded to correctly originally. Hence, it is necessary to design a text filter to filter the information independently. Information filtering ensures that large language models answer questions within their capabilities as much as possible, reducing the possibility of generating illusory facts.