MultiModal Robustness benchmark (MMR)
收藏资源简介:
MultiModal Robustness benchmark (MMR)是由北京人工智能研究院创建的一个多模态鲁棒性基准数据集,包含600条精心标注的正负问题对,涵盖12个类别。该数据集旨在评估多模态大型语言模型(MLLMs)在视觉理解和推理方面的能力,特别是在面对引导性问题时的鲁棒性。数据集通过提取图像中的视觉信息,如文本、对象、属性等,生成正负样本,以增强模型的理解和鲁棒性。MMR数据集的应用领域主要集中在提升MLLMs在复杂视觉场景下的准确性和稳定性,解决模型在面对误导性问题时可能出现的错误回答问题。
The MultiModal Robustness Benchmark (MMR) is a multimodal robustness benchmark dataset developed by the Beijing Academy of Artificial Intelligence. It contains 600 carefully annotated positive and negative question pairs across 12 distinct categories. This dataset is designed to evaluate the visual understanding and reasoning capabilities of multimodal large language models (MLLMs), with a particular focus on their robustness against leading questions. It generates positive and negative samples by extracting visual information such as text, objects, and attributes from images to enhance model understanding and robustness. The primary applications of the MMR dataset focus on improving the accuracy and stability of MLLMs in complex visual scenarios, and addressing the issue of erroneous responses generated by models when faced with misleading questions.




