Dataset

patent-classification

A sample of US patent titles, abstracts and CPC classification labels built for multi-class patent classification, tens of thousands of text records in Parquet. Useful as training material for mapping innovation activity onto technology categories over time.

Open on Hugging Face

Data preview

Subset

Loading preview...

Dataset card

Patent Classification is a public-records text dataset of roughly 35,000 English-language patents and their abstracts, distributed by the ccdv publisher. It frames patent documents as a multi-class text classification problem with nine coarse category labels across train, validation, and test splits.

What this dataset contains

The corpus holds about 35,000 U.S. patent records sampled from the BIGPATENT summarization collection, with abstracts and longer document bodies exceeding 512 tokens in many cases. It is distributed as parquet files through the datasets, pandas, and mlcroissant libraries and has been downloaded roughly 1,700 times. The license is not explicitly stated in the source listing.

Structure and fields

Each row exposes a `text` field containing the patent or abstract content as a string and a `label` field drawn from nine unbalanced category names such as "Human Necessities" and "Performing Operations; Transporting". Rows are partitioned into train (about 25,000), validation (about 5,000), and test (about 5,000) splits, totaling 70,000 rows. The category distribution is unbalanced by design, which is relevant when modelling.

How analysts use it

The dataset is positioned for training and benchmarking long-context text classifiers, and the documentation shows compatibility with the run_glue.py script using a 512-token maximum sequence length and a base RoBERTa model. Investment-fund use cases are not directly supported by the available facts, since the data are intended for NLP research rather than market or company-level signals. Readers seeking patent intelligence for idea-generation, prior-art screening, or competitive mapping should look to a dedicated patent-analytics product instead.

Documentation & references 1