MNBVC

Why we included this project

Any team building a Chinese-language model runs into the same wall: usable plain text is out there, but it is scattered, noisy, and other corpus projects have already picked it over. MNBVC goes after that gap at real scale, maintaining a steadily growing collection of Chinese text pulled from the internet, including news, novels, forums, wiki pages, subtitles, poetry, product listings, jokes, and chat logs, rough-cleaned into txt, json, jsonl, and parquet formats. The project is deliberately un-curated. The maintainers skip strict indexing and copyright review to keep downloads flowing, so treat it as raw material with source URLs and page screenshots recorded per folder, not as a polished benchmark. The corpus currently sits at roughly 60TB and is still growing toward its 253T target. Alongside the data, the project ships practical tooling for the ingest side, including charset detection, deduplication, format checking, and cleaning scripts, which most teams will end up reusing in their own pipelines. If you are pretraining or fine-tuning a Chinese LLM and want honest raw volume plus the plumbing to clean it, this is a good place to start.