USPTO AI Patent Dataset (AIPD)

The Artificial Intelligence Patent Dataset (AIPD) is a ready-made labelling of the US patent record, produced by the Office of the Chief Economist at the US Patent and Trademark Office. Rather than full text, it supplies a classification layer. For each of roughly 13.2 million granted patents and pre-grant publications published between 1976 and 2023, machine learning models predict whether the document involves artificial intelligence and, if so, which of eight component technologies it touches (machine learning, natural language processing, computer vision, speech, knowledge processing, AI hardware, evolutionary computation, and planning and control). The 2023 update rebuilt the models using BERT for Patents in place of the earlier Word2Vec embeddings, and the USPTO asks that users cite the accompanying paper.

You download it as flat tables in Stata (.dta) and CSV format from the USPTO Open Data Portal, where links now mint a temporary signed URL, or from the Open Data Portal bulk directory. The main predictions file is under a gigabyte, so the whole thing fits comfortably on a laptop. There is no query API, so you join the predictions to patent numbers yourself.

For RAG and AI work the value is as a filter and an enrichment layer, not a text corpus. Join it to a full-text source such as Google Patents Public Datasets, PatentsView, or the HUPD collection to pull only AI-related documents into an index, or to tag retrieved patents by AI technology when reranking. It is best for building a focused AI prior-art assistant or for studying trends in AI invention.

The licence position is unusually clean. The dataset carries a Creative Commons Public Domain Mark 1.0, so commercial use is permitted with no attribution or share-alike obligation, though a citation is requested as a courtesy. The one caveat is currency, because coverage ends in 2023 and will not reflect recent filings.