ChemPatents — Google Patents CPC C07 Snapshot (2000–2026)

A snapshot of 2,670,764 patent publications from Google Patents under CPC class C07 (organic chemistry), with publication dates from 2000-01-01 to 2026-05-08.

Records are family-deduplicated (one row per patent family; the deduplication column is Family for every row). Every publication_number in the file is unique.

Files

Path	Format	Size
data/chempatents_c07_2000_2026.parquet	Parquet (zstd-9)	326 MB
raw/google_patents_c07_current_snapshot_2000_2026.csv	Original CSV	1.75 GB

Both hold identical content. The Parquet file is the one the dataset viewer and load_dataset use; the CSV is kept for provenance. Row counts were verified to match (2,670,764) with two independent parsers.

Usage

from datasets import load_dataset

ds = load_dataset("EscheWang/ChemPatents", split="train")
print(ds[0])

Or read the Parquet directly — it is small enough to fit in memory:

import pandas as pd
df = pd.read_parquet("hf://datasets/EscheWang/ChemPatents/data/chempatents_c07_2000_2026.parquet")

Lazily, without materialising the whole table:

import polars as pl
lf = pl.scan_parquet("hf://datasets/EscheWang/ChemPatents/data/chempatents_c07_2000_2026.parquet")
lf.filter(pl.col("published_date").dt.year() == 2024).select("title").head().collect()

Schema

16 columns. No nulls anywhere — missing text values are the empty string "", as in the source CSV.

Column	Type	Description
publication_number	string	Publication ID, e.g. US10035800B2. Unique across the dataset.
title	string	Patent title. May be truncated with … in the source.
metadata	string	Unparsed blob: jurisdiction codes, publication number, inventor, assignee.
active_countries	string	;-separated country codes where the family is active. Empty for 9.6% of rows.
dates	string	Unparsed date line, e.g. Priority 1998-08-07 • Filed 1999-08-06 • Published 2001-02-06.
published_date	date32	Parsed publication date.
abstract_snippet	string	Search-result snippet: metadata + dates + a truncated abstract, newline-separated. Not a clean abstract.
patent_link	string	Google Patents URL.
pdf_link	string	PDF URL. Empty for 33.2% of rows.
interval_start	date32	Start of the date window used by the scraping query.
interval_end	date32	End of that window.
page	int64	Result page within the query.
result_rank	int64	Rank within the page.
absolute_page_rank	int64	Rank across the query's result set.
query	string	The Google Patents query that returned this row, e.g. CPC=C07D/low; country=CN.
deduplication	string	Always Family.

The last six columns describe how the row was collected, not the patent itself. Drop them for most modelling uses.

Composition

CPC subclass, by the query that retrieved each row (208 distinct query strings; grouped here):

Query family	Rows
C07 (broad)	1,508,364
C07D heterocyclic	411,516
C07K peptides	355,254
C07C acyclic/carbocyclic	236,388
C07F organo-element	53,381
C07H sugars/nucleosides	42,163
C07B general processes	38,201
C07J steroids	10,632
C07G unknown-constitution	993

Top jurisdictions by publication-number prefix: US 446,410 · CN 354,965 · EP 317,123 · WO 254,512 · JP 224,787 · KR 146,225 · AU 124,531 · DE 86,315 · CA 76,893 · TW 53,476.

⚠️ Known limitation: uneven year coverage

This is not a complete census of C07 publications, and coverage is very uneven across time. Rows per publication year:

Years	Rows/year
2000–2014	~93k–140k	dense
2015	78,823	thinning
2016–2022	5,524–25,031	sparse — badly under-covered
2023–2025	94,553–220,854	dense
2026	51,863	partial year (to 2026-05-08)

The 2016–2022 window has roughly 5–20× fewer records per year than the periods around it. This reflects gaps in how the snapshot was collected, not a real collapse in patenting activity. Do not use this dataset for time-series or trend analysis without accounting for it, and treat absolute counts in that window as a lower bound.

Other caveats:

• abstract_snippet is a search-result snippet, not a full abstract — it is prefixed with the metadata and date lines and the abstract text itself is often truncated with ….
• title is likewise truncated for long titles.
• No claims, no full text, no chemical structures (no SMILES/InChI). This is bibliographic metadata.
• Rows are retrieval records: page, result_rank, and query reflect a scraping process and carry no information about the patent.

Provenance

Scraped from Google Patents search results by CPC code over date windows, then deduplicated by patent family. Converted from the original CSV to Parquet with PyArrow (zstd level 9); the conversion is lossless and row counts were cross-checked with Python's stdlib csv reader.

License

The dataset card and the compilation are released under Apache-2.0. The underlying bibliographic records originate from Google Patents; patent documents themselves are public records, but check Google Patents' terms before redistributing derived products.