Usenet Corpus spans 103B tokens, 1980-2013

This is a serious corpus release, not just a novelty archive: the value is the combination of scale, time span, and documented cleaning, which makes it usable for long-context and historical language work.

• –Deduplication, PII redaction, binary filtering, and Message-ID hashing make the corpus much more training-friendly than raw Usenet dumps.
• –The 33-year arc is the standout feature: it captures pre-web, peak-web, and post-peak online discourse in one dataset.
• –96.6% English with 100+ other languages, especially in `soc.*`, gives it enough multilingual texture to matter without turning it into a generic multilingual dump.
• –The main caveat is provenance and bias: Usenet overrepresents technical, academic, and heavily quoted discussion, so it is best as a domain-adaptation source rather than a clean proxy for general internet language.