Scholarlink

The lack of specialized large language models (LLMs) for breast cancer limits their clinical adoption. To address this, we present Breast-CRAG, a 7B-parameter retrieval-augmented LLM specifically optimized for breast cancer applications. Our approach combines a fine-tuned generator (trained on a curated 268K-dialogue dataset) with a specialized retriever (leveraging a 1M-chunk knowledge base) to enhance response quality. Evaluations across four dialogue and two exam datasets demonstrate that Breast-CRAG outperforms comparable open-source models and achieves competitive performance with GPT-4o. Ablation studies confirm the contributions of both components. With its strong performance on clinical queries, Breast-CRAG represents a promising tool for breast cancer care and research.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Breast-CRAG: A Breast Cancer Large Language Model Leveraging Retrieval-Augmented Generation

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Zikang Chen,
  • Qinchuan Wang,
  • Jinyan Liu,
  • Yaoqian Sun,
  • Heming Zheng,
  • Haomin Li,
  • Huilong Duan,
  • Xudong Lu

摘要

The lack of specialized large language models (LLMs) for breast cancer limits their clinical adoption. To address this, we present Breast-CRAG, a 7B-parameter retrieval-augmented LLM specifically optimized for breast cancer applications. Our approach combines a fine-tuned generator (trained on a curated 268K-dialogue dataset) with a specialized retriever (leveraging a 1M-chunk knowledge base) to enhance response quality. Evaluations across four dialogue and two exam datasets demonstrate that Breast-CRAG outperforms comparable open-source models and achieves competitive performance with GPT-4o. Ablation studies confirm the contributions of both components. With its strong performance on clinical queries, Breast-CRAG represents a promising tool for breast cancer care and research.