COMPACTDS
收藏资源简介:
COMPACTDS是一个由高质量、多样化来源构建的3800亿单词的数据存储库,旨在与预训练数据的广度相匹配,同时保持可访问性。该数据集包括来自网络爬虫、精选的数学内容、学术论文和教科书的多样化数据。COMPACTDS的设计基于两个关键见解:一是可以在不影响覆盖范围的情况下过滤掉低质量的网络文本,从而得到一个更小但具有代表性的数据集;二是结合内存中的近似最近邻(ANN)检索和磁盘上的精确搜索,可以在单个节点上实现毫秒级的检索。
COMPACTDS is a 380-billion-word data repository constructed from high-quality, diverse sources, designed to match the breadth of pre-training data while remaining accessible. This dataset encompasses diverse data sourced from web crawls, curated mathematical content, academic papers, and textbooks. COMPACTDS is built upon two key insights: first, low-quality web text can be filtered out without compromising coverage, resulting in a smaller yet representative dataset; second, combining in-memory approximate nearest neighbor (ANN) retrieval with exact search on disk enables millisecond-level retrieval on a single node.




