illeatmyhat/uspto-trademarks

USPTO Trademark Case Files — analysis-ready Parquet

Complete USPTO trademark application/registration corpus, rebuilt from the official bulk XML products and normalized into relational Parquet tables.

• Sources: USPTO Open Data Portal products TRTYRAP (Trademark Full Text XML Data, No Images — Annual Applications; backfile from 1884) and TRTDXFAP (same, Daily Applications), US Trademark Applications v2.0 DTD.
• Data current through: 2026-07-06 (annual snapshot cutoff 2025-12-31 + daily replay).
• License: United States government work — public domain.
• Schema lineage: table/column names follow the USPTO Office of the Chief Economist (OCE) Trademark Case Files Dataset conventions, so existing research code ports directly. correspondent is an extension beyond OCE.

Querying (DuckDB, no download needed)

SELECT filing_dt, mark_id_char
FROM 'hf://datasets/illeatmyhat/uspto-trademarks/data/case_file/*.parquet'
WHERE serial_no = '75930757';

Tables

table	rows	grain
case_file	14,180,757	1 per case file
owner	32,438,405	1:M per serial_no
classification	16,843,367	1:M per serial_no
intl_class	17,010,081	1:M per serial_no
us_class	60,306,512	1:M per serial_no
statement	29,366,352	1:M per serial_no
event	246,827,790	1:M per serial_no
design_search	8,825,462	1:M per serial_no
prior_mark	2,571,981	1:M per serial_no
foreign_app	700,229	1:M per serial_no
madrid_intl_file	198,910	1:M per serial_no
owner_name_change	259,631	1:M per serial_no
correspondent	12,745,506	1 per case file

serial_no (8-char zero-padded string) joins everything. Row counts, per-column null rates and build lineage are published under profile/. USPTO's code tables (case status, prosecution-history event, statement type, party, legal entity, mark drawing) ship under lookups/, with per-table source URL and retrieval date in lookups/_sources.json.

Research & reproducibility

This corpus was assembled to study, quantitatively, the well-documented post-2015 surge in coined trademark filings tied to online-marketplace brand-gating (e.g. Amazon Brand Registry) — the phenomenon that motivated the Trademark Modernization Act of 2020. Dimensions of interest, all computable directly from the tables above: filing volume by year, owner geography, and marketplace-gated class (9/11/18/20/21/25/28); coined single-token marks; attorney-of-record filing concentration; the marks-per-owner ratio that distinguishes ongoing law practices from high-churn filing operations; goods/services identification length; and correspondent-address-level aggregation of filing operations.

A complete, runnable methodology — DuckDB queries plus a versioned scoring module that composes these signals into a ranked table — is in the open-source pipeline: docs/ANALYSIS.md. Every figure is a factual measurement of public records and reproducible in minutes. Results are descriptive patterns, not determinations about any filer; see the document's "Reading the results" note.

Update semantics & limitations

• Latest state only. Each case file appears once, in its most recent known state (annual baseline + daily replay, latest-wins). Superseded field values are not retained; consumers wanting version archaeology should rebuild from USPTO dailies themselves.
• Owner names are not normalized; address country codes contain invalid values — documented, not cleaned.
• Pre-1980s records are sparse/backfilled by USPTO.
• Every row carries file_dt_source/file_name_source lineage back to the exact USPTO bulk file that produced it.

Built 2026-07-09 by the USPTO Trademark Radar pipeline.