moltbook-factcheck-dose-response
收藏资源简介:
Moltbook Factcheck Dose-Response Experiment 是一个多智能体社交模拟数据集,用于测量不同比例的事实与阴谋内容对AI智能体在类似Reddit平台上的行为影响。数据集来源于一个剂量-响应实验,平台为Moltbook(一个为AI智能体设计的类似Reddit的社交网络),研究层为CivicLens(用于控制多智能体研究的实验基础设施)。实验使用LLM驱动的智能体(每次运行10个智能体加2个系统智能体),每次运行约1小时。数据集包含不同剂量水平(f0到f5)的25个世界帖子,比例从纯阴谋内容(f0)到最高比例的事实内容(f5)。帖子来源于真实的事实核查声明(Google Fact Check Explorer,2025年8月后),并通过Claude转换为Reddit风格的帖子对。数据集包含多个JSONL文件(posts.jsonl、comments.jsonl、treatments.jsonl、activity.jsonl、agents.jsonl)以及topic-mapping.json和metadata.json。总计175个帖子、177条评论和778个活动事件。适用于研究剂量-响应关系、智能体行为变化、信息级联动态和情感分析等任务。
The Moltbook Factcheck Dose-Response Experiment is a multi-agent social simulation dataset designed to quantify the impact of varying ratios of factual to conspiratorial content on AI Agents operating on a Reddit-like social platform. The dataset is derived from a dose-response experiment conducted on Moltbook—a Reddit-style social network specifically built for AI Agents—and utilizing CivicLens, an experimental infrastructure for controlling multi-agent research. The experiment uses LLM-powered Agents, with 10 Agents plus 2 system Agents per experimental run, and each run lasts approximately one hour. The dataset includes 25 world posts across different dosage levels (f0 to f5), with content ratios ranging from purely conspiratorial material (f0) to the highest proportion of factual content (f5). The posts are sourced from authentic fact-check claims indexed in Google Fact Check Explorer (post-August 2025), and converted into Reddit-style post pairs using Claude. The dataset contains multiple JSONL files: posts.jsonl, comments.jsonl, treatments.jsonl, activity.jsonl, and agents.jsonl, alongside topic-mapping.json and metadata.json. In total, the dataset encompasses 175 posts, 177 comments, and 778 activity events. This dataset is suitable for research tasks including dose-response relationship analysis, AI Agent behavioral changes, information cascade dynamics, and sentiment analysis.



