ulab-ai/swm-bench
收藏资源简介:
SWM-Bench是Social World Models(SWM)的基准数据集,用于预测市场集体信念(即价格)如何响应新闻变化,并确定哪些新闻驱动了这种变化。该数据集聚合了Polymarket和Kalshi市场(从2022年12月到2026年1月)的数据,并包含每次信念更新前可用的新闻,采样侧重于波动显著(新闻驱动)的变动。数据集包含原始价格序列和爬取的新闻(未处理),以及由Qwen3.5-397B和Qwen3-32B后验归因器标注的处理记录,这些标注数据是主要数据集。每个记录包括市场ID、事件ID、问题描述、类别、价格历史、候选新闻、归因分数(表示新闻对价格变动的责任程度)、目标价格、未来价格序列和标准化变动幅度。数据集用于训练和评估社会世界模型,支持预测市场分析和新闻归因任务。
SWM-Bench is the benchmark for Social World Models (SWM): predicting how a prediction markets collective belief (its price) shifts in response to news, and which news drives the shift. It aggregates Polymarket and Kalshi markets (Dec 2022 – Jan 2026) with the news available before each belief update, sampled toward volatility-significant (news-driven) moves. The dataset includes raw price series and crawled news (unprocessed), as well as processed records labeled by the Qwen3.5-397B and Qwen3-32B posterior attributors, which form the main dataset. Each record contains market ID, event ID, question, description, categories, price history, candidate news, attributions (with scores indicating responsibility for price moves), target price, future price sequence, and standardized move magnitude. The dataset is designed for training and evaluating social world models, supporting prediction market analysis and news attribution tasks.




