MMDS-Bench
收藏资源简介:
MMDS-Bench是一个专为社交媒体多模态动态立场分类设计的诊断基准,由武汉大学与上海创新研究院联合构建。该数据集包含3,482个多模态父-回复实例,覆盖七类立场标签,并额外设有800个实例的推理诊断子集。数据源自X平台,通过多领域关键词收集,经严格筛选与人工标注,定义了多模态融合、父消息框架等五类挑战因子。该基准旨在评估多模态大语言模型在复杂交互场景下的立场推理能力,尤其关注跨模态信号理解与关系推断的协同挑战。
MMDS-Bench is a diagnostic benchmark specifically designed for multimodal dynamic stance classification on social media, jointly developed by Wuhan University and Shanghai Institute of Innovation. This dataset includes 3,482 multimodal parent-reply instances covering seven stance categories, and additionally features an inference diagnostic subset of 800 instances. The data is sourced from X platform, collected through multi-domain keywords, and subjected to strict screening and manual annotation. The dataset defines five challenge factors including multimodal fusion and parent message framing. This benchmark aims to evaluate the stance reasoning ability of multimodal large language models in complex interactive scenarios, with a particular focus on the collaborative challenges of cross-modal signal understanding and relational inference.
MMDS-Bench 数据集详情
MMDS-Bench 是一个用于评估多模态大语言模型在社交媒体互动中动态立场识别能力的基准数据集。
核心信息
- 数据集名称:MMDS-Bench
- 所属领域:多模态大语言模型评估、社交媒体分析、立场检测
- 核心任务:动态立场(Dynamic Stance)识别,即在社交媒体互动情境中,评估模型对立场变化的理解与推理能力
- 数据特性:多模态(Multimodal),涵盖文本及可能的图像等非文本信息;动态性(Dynamic),关注互动过程中立场的演变而非静态判断
适用场景
该基准主要用于检验多模态大语言模型在复杂社交互动语境下的立场分析能力,适用于模型性能评测、立场检测算法研究及社交媒体舆情分析等相关任务。




