jeremycochoy/gift-pretrain-full-4096

gift-pretrain-full-4096是一个用于时间序列预测的预训练基准数据集，是gift-pretrain-small-4096的完整版本。它基于Salesforce/GiftEvalPretrain数据集，将每个arrow文件中的每个时间序列裁剪为长度为4096的非重叠窗口，并进行全局随机打乱。数据集包含6,376个源arrow文件，覆盖152个子数据集，生成42,571,692个窗口，总大小约619 GB。数据格式为Parquet分片，使用zstd压缩。数据集设计用于时间序列预测任务，支持预训练和评估。gift-pretrain-full-4096 is a pretraining benchmark dataset for time-series forecasting, serving as the full counterpart to gift-pretrain-small-4096. It is derived from the Salesforce/GiftEvalPretrain dataset, where every series from every arrow file is cropped into non-overlapping 4096-point windows and globally shuffled. The dataset includes 6,376 source arrow files across 152 sub-datasets, resulting in 42,571,692 windows with a total size of approximately 619 GB. Data is stored in Parquet shards with zstd compression. It is designed for time-series forecasting tasks, supporting pretraining and evaluation.