WebText

WebText is an internal OpenAI corpus created by scraping web pages with emphasis on document quality. The authors scraped all outbound links from Reddit which received at least 3 karma. The authors used the approach as a heuristic indicator for whether other users found the link interesting, educational, or just funny.

WebText 是 OpenAI 开发的内部语料库，其构建方式为抓取网页内容并着重筛选高质量文档。该数据集的构建者抓取了 Reddit 平台上获得至少 3 个 karma（Reddit社区用户积分）的全部出站链接，并将该积分阈值作为启发式判定指标，用以判断其他用户是否认为该链接内容有趣、具备教育价值或是单纯诙谐有趣。