MedHallTune
收藏资源简介:
MedHallTune是由香港中文大学等机构研发的大型数据集,包含超过10万张医学图像和100万条指令对,既有幻觉样本也有非幻觉样本,专门针对医疗应用设计。数据集内容丰富,涵盖了大量的医学图像和指令,旨在帮助模型更好地理解和处理医学场景中的幻觉问题,提高其在实际医疗应用中的可靠性和准确性。
MedHallTune is a large-scale dataset developed by The Chinese University of Hong Kong and other institutions. It contains over 100,000 medical images and 1 million instruction pairs, covering both hallucinatory and non-hallucinatory samples, and is specifically tailored for medical applications. Boasting rich content with a vast collection of medical images and instructions, this dataset aims to assist models in better understanding and addressing hallucination issues in medical scenarios, thereby enhancing their reliability and accuracy in real-world medical applications.
MedHallTune 数据集概述
数据集简介
MedHallTune是一个用于减轻视觉语言模型中的医学幻觉的基准和指令调整数据集。
数据集发布
- 训练和评估的MedHallTune数据集以及模型权重即将发布。
- 数据集已在arXiv上可用。
数据集下载
- 数据集可在Huggingface Hub上获取。
引用信息
若MedHallTune对您的研究有用或相关,请通过以下引用认可我们的贡献:
bibtex @misc{yan2025medhalltune, title={MedHallTune: An Instruction-Tuning Benchmark for Mitigating Medical Hallucination in Vision-Language Models}, author={Qiao Yan and Yuchen Yuan and Xiaowei Hu and Yihan Wang and Jiaqi Xu and Jinpeng Li and Chi-Wing Fu and Pheng-Ann Heng}, year={2025}, eprint={2502.20780}, archivePrefix={arXiv}, primaryClass={cs.AI}, url={https://arxiv.org/abs/2502.20780}, }

- 1MedHallTune: An Instruction-Tuning Benchmark for Mitigating Medical Hallucination in Vision-Language Models香港中文大学计算机科学与工程系,香港中文大学医学智能与XR研究所,上海人工智能实验室 · 2025年



