HartvigsenGroup/BEDTime

BEDTime是一个包含10175个时间序列及其对应的自然语言描述的数据集，旨在评估大型语言和视觉语言模型在时间序列到文本生成任务上的表现。数据集由四个来源组成：TRUCE-Stock（真实股票数据），TRUCE-Synthetic（展示特定模式的合成数据），SUSHI-Tiny（具有趋势、周期性和波动类型的合成数据），以及TaxoSynth（基于属性分类和模板生成文本的合成数据）。

BEDTime is a dataset containing 10,175 time series along with their corresponding natural language descriptions, designed to evaluate large language and vision-language models on the task of time series to text generation. The dataset is composed of four sources: TRUCE-Stock (real-world stock data), TRUCE-Synthetic (synthetic data exhibiting specific patterns), SUSHI-Tiny (synthetic data with trends, periodicities, and fluctuation types), and TaxoSynth (synthetic data annotated with attribute-based taxonomy and template-based text generation).