Dataset

us_patent

Public Records & Filings Download 1 source

A reference collection drawn from the U.S. Patent Phrase to Phrase Matching Kaggle competition, containing additional details that are still required for full documentation.

Open on Hugging Face

Data preview

Subset

Loading preview...

Dataset card

The us_patent dataset is a text corpus from the U.S. Patent Phrase-to-Phrase Matching benchmark, distributed in parquet format via the Hugging Face datasets library. It maps short anchor phrases to longer target phrases within patent classification codes and assigns a similarity score to each pair.

What this dataset contains

The dataset consists of roughly 36,500 phrase-to-phrase pairs drawn from U.S. patent documents, each annotated with a similarity score. Pairs are assigned to a CPC classification context such as A47, allowing analysts to subset the corpus by technology area. Two representative examples pair the anchor "abatement" with targets like "abatement of pollution" (score 0.5) and "act of abating" (score 0.75).

Structure and fields

Records are split into train and test partitions and stored as parquet with five columns: id (string), anchor (string), target (string), context (string CPC code), and score (float32). The id field is a hash identifier, anchor and target hold short text phrases, and context indicates the Cooperative Patent Classification section. Score is the human-assigned or model-derived similarity label between the two phrases.

How analysts use it

Because the file is a labeled text-matching benchmark, the most direct application is training or evaluating natural language models that score semantic similarity between patent terms, optionally filtered by context code. Investment workflows are limited: the data does not include filing dates, assignees, geographies or citation links, so it cannot itself support patent activity, innovation intensity or freedom-to-operate screens without enrichment from external patent records.

Documentation & references 1