Skip to content

Dataset · HUPD

Harvard USPTO patent application corpus

HUPD/hupd

Over 4.5M US patent application drafts (2004-2018) for patentability analysis, classification, and summarization.

What's inside

It contains over 4.5M English patent documents submitted to the United States Patent and Trademark Office (USPTO) from 2004 to 2018. Unlike alternatives, the dataset contains the original application versions from inventors rather than final approved texts. The corpus was prepared by Harvard researchers for multi-purpose NLP tasks.

What you can do with it

  • Training models to predict patentability of an application at submission
  • Fine-tuning models for summarization and classification of legal and technical texts
  • Pre-training language models in the patent domain (fill-mask)
Tasks
Fill mask, Summarization, Text classification, Token classification
Languages
English
Downloads this month
5.3K
Likes
45

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("HUPD/hupd")