BIGPATENT

BIGPATENT is a large-scale patent text corpus assembled by researchers at Northeastern University (Eva Sharma, Chen Li, and Lu Wang) and introduced in a 2019 paper on abstractive and coherent summarisation. It gathers roughly 1.3 million US utility patents filed to the USPTO between 1971 and 2018, drawn from the Google Patents Public Datasets via BigQuery, and pairs each patent's full description with its human-written abstract. The abstract serves as the gold-standard summary and the description as the input, which makes the corpus a ready-made training set for long-document summarisation rather than a general patent search tool. Records are organised across the nine Cooperative Patent Classification (CPC) categories, so you can slice the data by technical field.

Access is straightforward. The dataset is hosted on Hugging Face and distributed as Parquet, so you can load it with the standard datasets library, stream it instead of downloading the whole thing, and filter by split or CPC class. It ships as predefined train, validation, and test splits, which suits model training and benchmarking out of the box.

For RAG and AI work, BIGPATENT is best treated as a clean, long-form English corpus of patent prose: useful for fine-tuning or evaluating summarisation, for patent classification, and as a retrieval corpus for IP-focused question answering. Patent text is dense and formulaic, so retrieval improves with chunking that respects section structure.

The licence is CC BY 4.0: commercial use is permitted and there is no share-alike obligation, but you must credit the dataset. Its main limits are scope and freshness. It covers US utility patents only, stops in 2018 rather than updating live, and is shaped around the summarisation task. Among our patent sources, reach for HUPD when you want examination decisions and application-level structure, PatentsView or Google Patents Public Datasets for the broadest live bibliographic coverage, and BIGPATENT when you specifically need patent body text paired with abstract summaries at scale.