pietrolesci/pile-deduped-stats

该文件夹包含研究中每个模型大小的模型评估（或“统计”），并且这些统计数据已经在序列级别进行了聚合。基于“原始”版本，其中包含令牌级别的统计数据（`pietrolesci/pythia-deduped-stats-raw`）。

This dataset contains model evaluation statistics for various model sizes, already aggregated at the sequence level. The dataset features include sequence index, step, support sequence, entropy sequence, cumulative sequence, and average rank, with data types of integers and floats. The dataset is divided into multiple parts, each with different byte sizes and example counts. The download size of the dataset is 572947240 bytes, and the actual size is 651225600 bytes. The dataset configuration includes the default configuration, with each configuration corresponding to different data file paths. The dataset is based on the raw version (token level) data for sequence-level statistics.