We follow the FineWeb-Edu pattern: sample a small, stratified set of documents across the five most-common languages, label each with an LLM, then train a lightweight multilingual text classifier on those labels and apply it at scale.
-The classifier is a fine-tuned XLM-RoBERTa, trained on roughly 3,500 LLM-labeled examples. Two independent classifiers are trained, one per dimension. Code is open: see the classification scripts on GitHub.
+The classifier is a fine-tuned XLM-RoBERTa, trained on roughly 3,500 LLM-labeled examples. Two independent classifiers are trained, one per dimension. Code is open: see the classification scripts on GitHub.
docx-corpus is the largest open corpus of classified Word documents on the public web. It contains 736,000+ real .docx files collected from Common Crawl, validated, deduplicated, and labeled by document type and topic. It is the first large-scale research dataset for DOCX, the world's most-used document creation format, which has historically been absent from document AI work built on scanned images and PDFs.
-Each document has a content hash, file size, detected language, and a type and topic label from a fine-tuned XLM-RoBERTa text classifier. The corpus spans 46+ languages and uses a 10-type, 9-topic taxonomy. Available as a HuggingFace dataset, a REST API at api.docxcorp.us, and downloadable manifest files. Built by SuperDoc. Source on GitHub. Dataset is ODC-BY; source code is MIT.
+Each document has a content hash, file size, detected language, and a type and topic label from a fine-tuned XLM-RoBERTa text classifier. The corpus spans 46+ languages and uses a 10-type, 9-topic taxonomy. Available as a HuggingFace dataset, a REST API at api.docxcorp.us, and downloadable manifest files. Built by SuperDoc. Source on GitHub. Dataset is ODC-BY; source code is MIT.