Skip to content
DataForge

Patent Dataset Download: USPTO Bulk Data, PatentsView, and the Citation Graph

Published 2026-08-04 · DataForge team

Patents are the densest public record of who is inventing what: 9M+ US grants since 1976, each with inventors, owners (assignees), technology classifications, and citations to prior art forming one of the largest directed graphs in open data. All of it is free from the USPTO. The gap between “free” and “usable”, though, is wide enough that an entire research infrastructure (PatentsView) exists just to bridge it — and even PatentsView leaves real work on the table. Here’s the honest map.

The free sources

USPTO bulk products. Weekly full-text XML of every grant and application since 1976 (at bulkdata.uspto.gov), plus specialized files (maintenance fees, assignments/reassignments, litigation). Authoritative and enormous — parsing decades of evolving XML DTDs is a serious engineering project.

PatentsView. The USPTO-funded research database: pre-parsed relational tables (patents, inventors, assignees, CPC classes, citations) as bulk TSV downloads plus an API. This is the right starting point for almost everyone, and it’s what our dataset builds on:

import pandas as pd

# PatentsView bulk: granted patents table (~9M rows)
patents = pd.read_csv(
    "g_patent.tsv.zip",
    sep="\t",
    usecols=["patent_id", "patent_date", "patent_title", "patent_type"],
    parse_dates=["patent_date"],
)
print(patents[patents.patent_date >= "2024-01-01"].shape)

Google Patents Public Data. BigQuery datasets including worldwide bibliographic data — excellent if you live in BigQuery and can pay the query costs; awkward as a local corpus.

EPO/WIPO. PATSTAT (paid, the academic standard for global analysis) and free OPS APIs for European/international coverage.

Where the real work hides

Assignee disambiguation. The famous problem. “International Business Machines Corporation”, “IBM CORP”, “IBM” plus hundreds of subsidiaries must resolve to one company before any “top patenters” or company-innovation analysis is valid. PatentsView ships algorithmic disambiguation — good, not perfect, and not joined to any external company universe. Linking assignees to SEC filers, government contractors or hiring data is a separate entity-resolution project (that link is exactly what our Company360 master provides).

The citation graph’s scale. Patent-to-patent citations are hundreds of millions of edges. Graph metrics (originality, generality, forward-citation impact) need real data engineering — the full relationship tables in our M tier run 255,347,795 rows.

Citation semantics. Examiner-added vs applicant-cited references mean different things for knowledge-flow studies; truncation bias (recent patents haven’t had time to be cited) invalidates naive impact comparisons across cohorts. These are analytical traps, not parsing traps — no dataset fixes them for you, but a clean schema makes handling them tractable.

-- Forward-citation impact, cohort-controlled: top 2015 patents
SELECT p.patent_id,
       p.patent_title,
       COUNT(c.citing_patent_id) AS forward_cites
FROM patents p
JOIN citations c ON c.cited_patent_id = p.patent_id
WHERE p.patent_date BETWEEN '2015-01-01' AND '2015-12-31'
GROUP BY p.patent_id, p.patent_title
ORDER BY forward_cites DESC
LIMIT 20;

CPC classifications: the underrated analysis axis

Most patent analyses group by company; the sharper ones group by technology, and the Cooperative Patent Classification (CPC) is how. Every patent carries one or more CPC codes in a hierarchical taxonomy — G06N is machine-learning/AI computing arrangements, H01M is batteries, A61K is medicinal preparations — maintained jointly by the USPTO and EPO and updated as fields emerge (a dedicated Y02 branch tracks climate-mitigation technologies across all fields, and G06N 3/0475-era subdivisions now separate generative models). Three analyses that CPC unlocks and company-level grouping can’t: technology emergence curves — count grants per CPC subgroup per year and you get a defensible measure of where inventive activity is accelerating, years before market data exists; company technology fingerprints — the CPC distribution of a firm’s portfolio, whose year-over-year drift shows strategic pivots (watch an automaker’s fingerprint migrate from F02 combustion engines toward H01M cells); and white-space mapping — CPC pairs that rarely co-occur on the same patent but increasingly co-occur in the same company’s portfolio flag integration plays before products ship.

The practical warning: patents average ~3–4 CPC codes each, so CPC-level counting triple-counts unless you decide a convention (primary classification only, fractional counting, or full counting with disclosure). Our CPC table keeps the full assignment list with sequence order, so any of the three conventions is a GROUP BY away — and the datasheet states which convention our own summary stats use.

What people build with patent data

  • Competitive technology intelligence. What is company X patenting, in which CPC classes, accelerating or decelerating — visible ~18 months after filing.
  • Innovation economics. The Hall–Jaffe–Trajtenberg tradition: citations as knowledge flows, patent counts as innovation output. Still one of the most active empirical literatures.
  • M&A and IP due diligence. Portfolio mapping, citation-based quality screens, reassignment histories.
  • ML on patents. Classification (title/abstract → CPC), prior-art retrieval, generation evals. 9M titled, classified documents with a citation graph make an exceptional training corpus — note our AI-training license tier below.
  • Talent mapping. Inventor mobility between companies, from disambiguated inventor records.

One tooling note before the comparison: at this scale, columnar formats and columnar engines aren’t optional. The citation table alone, as CSV, is a >10GB parse every time you touch it; as Parquet with DuckDB or Polars, the forward-citation query above runs in seconds on a laptop. If you assemble from PatentsView TSVs yourself, converting to Parquet with explicit dtypes (patent IDs as strings — they contain letters for design/plant patents, and integer coercion silently corrupts them) is the first hour best spent. Our packages ship Parquet-first for exactly this reason.

Free sources vs. our packaged corpus

USPTO XML PatentsView DataForge dataset
Cost Free Free From $499
Parsing All yours Done Done
Core patents — 9M+ TSVs 9,075,421 patents, analysis-ready Parquet/CSV
Relationship tables — Separate large TSVs Citations, assignees, inventors, CPC — 255M rows, one schema (M tier)
QA / docs — Research docs Data dictionary, datasheet, QA report per package
AI training Unclear stack Research terms Explicit Tier-2 AI/ML training license (L tier, approval required)

Honest guidance: PatentsView bulk files are free and good — if you’re comfortable stitching multi-GB TSVs and their update cycle, start there. Our packaging buys you the assembled, QA’d corpus with documented conventions, and — uniquely — a clear commercial license including an explicit AI-training tier, which matters if your legal team has opinions about training on “research use” data. Sample first: free 500-row sample.

FAQ

Is patent data public domain? Patent documents are US government publications free of copyright concerns. What licenses actually cover (PatentsView’s terms, ours) is the compiled database and derived fields — disambiguation, normalization, packaging.

How current is the data? USPTO grants publish weekly; PatentsView releases quarterly; our editions are dated snapshots covering grants 1976–2024.

US only? Yes, currently — US grants. For global families, EPO’s PATSTAT is the standard complement; our schema keeps standard identifiers to make that join feasible.

How do I track patent ownership changes (reassignments)? The USPTO’s assignment database records transfers — sales, mergers, security interests — as a separate bulk product, and it’s underused because it’s messy: recording is voluntary-ish (incentivized, not perfectly enforced), one recordation can cover thousands of patents, and conveyance types need text parsing (“assignment” vs “security agreement” vs “merger” mean very different things — a bank taking patents as loan collateral is not an acquisition). Handled carefully, reassignment data powers some of the most interesting analyses in the field: technology-asset M&A archaeology, patent-troll supply chains (operating companies → NPEs), and collateralization as a distress signal.

Is patent text useful for LLM work, and what are the gotchas? Very — patents are long, technical, formally structured documents with human-assigned classification labels and a citation graph: excellent for retrieval benchmarks, classification fine-tuning and domain adaptation. Gotchas: OCR noise in pre-2001 grants (parsed from scanned images), legalese register that transfers poorly to general tasks, claims vs description being different linguistic animals (claims are single-sentence legal constructions hundreds of words long), and the license question — full text from USPTO is public record, but compiled corpora with derived fields carry database terms, which is why our L tier makes AI-training rights explicit instead of leaving them to vibes.

What’s the lag between invention and visibility in the data? Applications publish 18 months after filing; grants add years of examination on top. Patent data is a lagging indicator of R&D and a leading indicator of products. Calibrate expectations accordingly: you’re seeing 2024’s inventions in 2026’s grants — still years before most reach market.

Do I need the full 255M-row graph, or is the 9M-patent core enough? Depends entirely on whether your unit of analysis is the patent or the relationship. Portfolio counts, grant trends and company fingerprints run fine on the core table. Citation impact, inventor mobility and knowledge-flow work need the relationship tables — that’s what the M tier adds.

Applications or grants? Grants. Applications (including never-granted ones) are a separate USPTO corpus with different semantics — valuable, but not in this package.