ENTRAP-VL
收藏资源简介:
ENTRAP-VL是由印度多所研究机构联合创建的一个专门用于评估视觉-语言模型中双上下文牵引现象的数据集。该数据集包含1500个条目,分为文本牵引流和视觉牵引流,涵盖八个类别,通过手动精心构建以体现其分类学结构。其创建过程基于对项目关联性和真实性关系的双重轴分类,旨在探究模型在文本和视觉两种独立上下文影响下的行为偏差。该数据集主要应用于行为分析领域,旨在解决视觉-语言模型在辅助上下文影响下可能产生的非理性输出问题,为可信人工智能和跨模态决策解码研究提供结构化评估工具。
ENTRAP-VL is a dataset jointly developed by multiple research institutions across India, specifically tailored to evaluate the dual context priming phenomenon in vision-language models. It consists of 1500 entries, categorized into two streams: text priming stream and visual priming stream, spanning eight categories. The dataset is meticulously handcrafted to reflect its taxonomic framework. Its development adopts a dual-axis classification based on project relevance and authentic relationality, with the goal of exploring behavioral biases exhibited by models under the influence of two independent contexts: textual and visual. Primarily applied in the field of behavioral analysis, this dataset aims to resolve the problem of irrational outputs that vision-language models may produce when affected by auxiliary contexts, providing a structured evaluation tool for trustworthy artificial intelligence and cross-modal decision-making decoding research.





