Reproduce “MedFact: A Large-scale Chinese Dataset for Evidence-based Medical Fact-checking of LLM Responses” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Tong Chen, Zimu Wang, Yiyi Miao, Haoran Luo, Yuanfei Sun, Wei Wang, Zhengyong Jiang, Procheta Sen and 1 more

MEDFACT, the first large-scale, evidence-based Chinese dataset, was developed to fact-check medical content generated by Large Language Models (LLMs). Evaluations using MEDFACT show that current LLMs do not reach human performance in this nuanced task, particularly struggling with "Uncertain" and "Not Applicable" claims.

Key insights

  • Current LLMs, whether in-context or fine-tuned, do not surpass human performance on this medical fact-checking task, indicating its inherent complexity and the need for further specialization.
  • Fine-tuning smaller models on the specific MEDFACT dataset can lead to performance that exceeds some significantly larger, non-fine-tuned LLMs, highlighting the value of domain-specific adaptation.
  • LLMs consistently struggle with nuanced veracity categories like "Uncertain" and "Not Applicable," often due to challenges in evidence understanding, semantic containment, and medical synonymy.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.