Unison
收藏资源简介:
Unison是一个综合性基准数据集,包含2,169个高质量的统一任务样本,旨在评估统一多模态模型中的联合理解与生成能力。该数据集围绕四个核心任务构建:内部一致性(IC),衡量理解与生成之间的内部对齐;理解引导生成(UGG),评估模型利用理解来指导上下文生成的能力;生成引导理解(GGU),探究合成输出如何辅助理解任务;相互增强(ME),通过多轮迭代协同,使理解识别生成错误,反之亦然,实现相互优化。数据内容涵盖文本提示、问题对、图像以及基于区域的编辑操作,具体组织为四个主要目录:Internal_Consistency(包含556个生成提示、对应的是/否视觉问答和参考图像)、Und_Guided_Gen(包含543个基于图像的生成项目,涉及边界框/掩码和操作类型)、Gen_Guided_Und(进一步分为2D空间、3D空间和复杂关系三个子类别,共542个项目)以及Mutual_Enhancement(包含264行数据,对应528个样本和源图像)。数据集适用于多模态人工智能模型的基准测试,特别是在需要联合处理视觉与语言理解、生成及其交互的任务中。
Unison is a comprehensive benchmark dataset containing 2,169 high-quality unified task samples, designed to evaluate joint understanding and generation capabilities in unified multimodal models. The dataset is built around four core tasks: Internal Consistency (IC), which measures the internal alignment between understanding and generation; Understanding-Guided Generation (UGG), which assesses the models ability to leverage understanding to guide contextual generation; Generation-Guided Understanding (GGU), which explores how synthetic outputs can assist understanding tasks; and Mutual Enhancement (ME), which enables mutual optimization through multi-round iterative collaboration, where understanding identifies generation errors and vice versa. The data content includes text prompts, question pairs, images, and region-based editing operations, organized into four main directories: Internal_Consistency (containing 556 generation prompts, corresponding yes/no visual question-answering, and reference images), Und_Guided_Gen (containing 543 image-based generation projects involving bounding boxes/masks and operation types), Gen_Guided_Und (further divided into three subcategories: 2D spatial, 3D spatial, and complex relationships, totaling 542 projects), and Mutual_Enhancement (containing 264 rows of data, corresponding to 528 samples and source images). The dataset is suitable for benchmarking multimodal AI models, particularly in tasks requiring joint processing of visual and language understanding, generation, and their interactions.




