MARS
收藏资源简介:
MARS数据集由香港科技大学计算机科学与工程系创建,包含35.5万条标注数据,用于评估大型语言模型在处理分布变化时的元物理推理能力。该数据集通过ChatGPT从Wikitext和BookCorpus中提取事件,并对其可变组件进行抽象化和数值变异,以创建元物理推理状态。MARS数据集旨在解决语言模型在面对环境因素和自身或其他代理行动引发的分布变化时的推理挑战,适用于评估和提升语言模型在复杂环境中的推理和规划能力。
The MARS Dataset was developed by the Department of Computer Science and Engineering at the Hong Kong University of Science and Technology. It contains 355,000 annotated data samples, which are designed to evaluate the meta-physical reasoning capabilities of large language models (LLMs) when handling distribution shifts. This dataset extracts events from Wikitext and BookCorpus via ChatGPT, then abstracts their variable components and performs numerical mutations to create meta-physical reasoning states. The MARS Dataset aims to address the reasoning challenges faced by language models when encountering distribution shifts caused by environmental factors, their own actions or those of other AI agents, and is suitable for evaluating and enhancing the reasoning and planning abilities of language models in complex environments.

- MARS数据集首次发表,由加州大学伯克利分校的研究团队提出,旨在为多目标跟踪任务提供一个标准化的评估平台。
- MARS数据集首次应用于多目标跟踪算法的研究,成为该领域的重要基准数据集之一。
- MARS数据集被广泛应用于多个国际计算机视觉会议(如CVPR、ICCV)的论文中,进一步验证了其在多目标跟踪任务中的有效性。
- MARS数据集的扩展版本发布,增加了更多的视频序列和标注信息,以支持更复杂的多目标跟踪任务研究。
- MARS数据集被用于多个多目标跟踪挑战赛,推动了该领域算法性能的提升和创新。



