Seed42Lab/RefOI-TLHF

RefOI-TLHF数据集是一个针对指代表达式的标记级别人类反馈的伴侣数据集，旨在对视觉语言模型在指代表达式生成中的语用学能力进行评估。该数据集包含图像、掩码、框选图像、边界框坐标、类别标签、是否为COCO类别标识、同类干扰物数量、指代表达式的来源、注释、完整描述、最小信息跨度标记的token列表以及最小token长度与完整token长度的比例等信息，用于进行指代语言中的冗余性和信息性的细粒度语言分析。

The RefOI-TLHF dataset is a companion dataset focusing on token-level human feedback for referring expressions, aiming to evaluate the pragmatic ability of vision-language models in the generation of referring expressions. It contains image, mask, boxed image, bounding box coordinates, class labels, COCO category indicator, number of same-class distractors, source of the referring expression, notes, full description, a list of minimally informative token spans, and the ratio of the length of minimal tokens to the full token length, for fine-grained linguistic analysis of redundancy and informativeness in referring language.