Harvard USPTO Patent Dataset (HUPD)

The Harvard USPTO Patent Dataset (HUPD) is a large, cleanly structured corpus of US patent applications built specifically for machine learning and natural language processing (NLP), the field concerned with teaching computers to work with human language. It draws on filings to the US Patent and Trademark Office and, unlike most patent tools, is shaped for people training and evaluating models rather than for lawyers running prior-art searches.

Its structure is what sets it apart. Each record separates the parts of an application, title, abstract, claims, background, and summary, and, crucially, includes the examination decision, so every patent is labelled by whether it was accepted or rejected. That turns it into a ready-made dataset for the tasks its authors designed it around, such as predicting patentability from the text or studying how language differs between granted and rejected filings.

Getting started is about as easy as it gets. It is hosted on Hugging Face and distributed as `Parquet` and `JSON`, so you can load it with the standard `datasets` library in a couple of lines, stream it rather than downloading it whole, and filter by filing year or decision. That low friction makes it a sensible starting point for a focused US-patent RAG or a quick fine-tuning experiment.

The licence is CC BY 4.0: commercial use is fine as long as you credit HUPD, with no share-alike obligation. The main limits are scope and freshness. It covers US applications only, over a defined range of years rather than updating live, and patent prose is dense and formulaic, so retrieval improves a lot with careful chunking that respects the claim and section structure.

If you need the widest international coverage or live examination events, Google Patents Public Datasets is the broader (if heavier) choice. HUPD is the one to reach for when you want a clean, ML-ready US corpus you can load and start building with the same day.