SocialOmni
收藏资源简介:
SocialOmni是由厦门大学等机构联合构建的多模态社交互动基准数据集,旨在评估全模态大语言模型在动态对话中的社交交互能力。该数据集包含2,209个高质量多轮对话样本,涵盖15个领域,分为感知任务(2,000条多选问题)和生成任务(209条开放问答)。数据来源于严格筛选的公共视频平台CC-BY授权内容,通过自动转录和人工标注构建。其核心应用领域为测试模型在说话人识别、打断时机控制及自然打断生成三大维度的综合表现,解决现有基准对实时社交能力评估不足的问题。
SocialOmni is a multimodal social interaction benchmark dataset jointly constructed by Xiamen University and other institutions, aiming to evaluate the social interaction capabilities of full-modal large language models in dynamic conversations. This dataset includes 2,209 high-quality multi-turn dialogue samples spanning 15 domains, which are divided into two categories: perception tasks (2,000 multiple-choice questions) and generation tasks (209 open-ended questions). The data is sourced from strictly screened CC-BY licensed content on public video platforms, and constructed through automatic transcription and manual annotation. Its core application is to test the comprehensive performance of models across three dimensions: speaker recognition, interruption timing control, and natural interruption generation, addressing the issue that existing benchmarks fail to adequately evaluate real-time social interaction capabilities.
SocialOmni 数据集概述
数据集基本信息
- 数据集名称:SocialOmni
- 发布平台:Hugging Face
- 许可证:MIT License
许可证信息
- 许可证类型:MIT License
- 许可证链接:https://huggingface.co/datasets/alexisty/SocialOmni/blob/main/LICENSE




