ComplexDataLab/socsim26-sharedtask
收藏资源简介:
该数据集是COLM 2026研讨会关于基于LLM的社会模拟共享任务发布的扫掠数据,用于构建评估方法以分析社会模拟日志。它包含五个场景研究:11-20金钱请求游戏(beauty_contest,涉及层级推理和玩家角色)、10轮双玩家囚徒困境(iterated_pd,涉及框架和角色立场)、回声室极化研究(polarization,涉及曝光、拓扑和记忆)、观察到的规范研究(observed_norms,基于世界价值观调查的四个国家人口数据)以及角色表达研究(persona_expression,探索代理群体多样性的来源)。每个研究提供假设供评估,并允许用户基于场景和数据提出自己的问题。数据以压缩包形式提供,解压后包含设计文件、运行日志(如动作事件、探测事件和模型调用记录)和元数据。数据集使用CC BY 4.0许可证,适用于研究目的。
This dataset is the released sweep data for the shared task at the COLM 2026 Workshop on LLM-Based Social Simulations, aimed at building evaluation methods over social simulation logs. It includes five scenario studies: the 11-20 money request game (beauty_contest, involving level-k reasoning and player personas), a 10-round two-player prisoners dilemma (iterated_pd, involving framing and persona stances), echo chambers across exposure, topology, and memory (polarization), WVS value battery across four country populations (observed_norms), and where agent-population diversity comes from (persona_expression). Each study provides hypotheses for evaluation and allows users to pose and test their own questions grounded in the scenario and data. The data is provided as tarballs that extract to directories containing design files, run logs (e.g., action events, probe events, and model call records), and metadata. The dataset is released under CC BY 4.0 license for research use.




