The 20 Newsgroups dataset is a collection of ~20k newsgroup documents and is super popular for experimentation in text applications of various machine learning techniques. Without Usenet (and archives of Usenet) that probably wouldn't exist.