---
pretty_name: PATSTAT AI Complete Master, 1950-2026
language:
- en
license: other
task_categories:
- text-classification
tags:
- patents
- artificial-intelligence
- patstat
- bibliometrics
- innovation
- cpc
- ipc
size_categories:
- 1M<n<10M
configs:
- config_name: all_periods
  data_files:
  - split: train
    path: data/by_priority_period/*.parquet
- config_name: 1950s
  data_files:
  - split: train
    path: data/by_priority_period/patstat_ai_1950s.parquet
- config_name: 1960s
  data_files:
  - split: train
    path: data/by_priority_period/patstat_ai_1960s.parquet
- config_name: 1970s
  data_files:
  - split: train
    path: data/by_priority_period/patstat_ai_1970s.parquet
- config_name: 1980s
  data_files:
  - split: train
    path: data/by_priority_period/patstat_ai_1980s.parquet
- config_name: 1990s
  data_files:
  - split: train
    path: data/by_priority_period/patstat_ai_1990s.parquet
- config_name: 2000s
  data_files:
  - split: train
    path: data/by_priority_period/patstat_ai_2000s.parquet
- config_name: 2010s
  data_files:
  - split: train
    path: data/by_priority_period/patstat_ai_2010s.parquet
- config_name: 2020_2026
  data_files:
  - split: train
    path: data/by_priority_period/patstat_ai_2020_2026.parquet
---

# PATSTAT AI Complete Master, 1950-2026

## Dataset Summary

This research dataset contains 2,330,553 unique patent applications and 175 columns related to artificial-intelligence patents. It integrates a legacy AI patent master with newly collected PATSTAT Online records. Duplicate resolution uses `app_id`, and the legacy record takes precedence when an application appears in both sources.

The data are partitioned by `priority_year` to support selective loading. The `2020_2026` category includes priority years 2020 through 2026. Records outside the documented range, if any, are retained in explicit boundary or unknown-year categories.

## Repository Structure

```text
data/by_priority_period/  Parquet data partitions by priority-year category
documentation/            Technical report and dataset report
metadata/                 Schema, manifests, descriptive statistics, and QA reports
code/                     Reproducible integration and backfill scripts
```

## Data Sources and Version

- Primary database: PATSTAT Online 2026 Spring.
- Legacy component: 1,850,664 legacy-only applications.
- Overlap component: 186,715 applications present in both sources; legacy values take precedence.
- New component: 293,174 new-only applications.
- Final key: 2,330,553 unique `app_id` values; no duplicate application IDs.

## Categories

The main configuration loads all Parquet files. Period-specific configurations load a single priority-year category. These are storage and access categories, not machine-learning labels.

```python
from datasets import load_dataset

all_data = load_dataset("deep1003/PATSTAT-AI-Complete-Master-1950-2026", "all_periods")
recent = load_dataset("deep1003/PATSTAT-AI-Complete-Master-1950-2026", "2020_2026")
```

The repository is publicly accessible. Users remain responsible for complying with the terms and restrictions described below.

## Core Variables

Core identifiers and content fields include `app_id`, patent office, application number, priority year, title, abstract, applicant information, inventor information, applicant and inventor country codes, IPC codes, and CPC codes. The complete 175-column schema is available in `metadata/schema.json`; column-level completeness is available in `metadata/column_descriptive_statistics.csv`.

IPC codes are present for 2,326,676 applications (99.8336%). CPC codes are present for 1,850,180 applications (79.3880%). Missing CPC values can reflect the absence of a PATSTAT CPC relation and should not automatically be interpreted as collection failure.

## Country Completion

Country fields include observed and rule-based completed values. The completion logic uses applicant, inventor, family, name, and patent-office evidence. Users must inspect the method and provenance fields before treating a completed country as directly observed. Patent-office country is not equivalent to inventor nationality or residence.

## Intended Uses

The dataset is intended for patent landscaping, bibliometric research, technology-trend analysis, innovation studies, and reproducible methodological evaluation. It is not intended for legal-status determination, individual profiling, automated decisions about people, or inference of nationality from names.

## Limitations

- Inventor and applicant country values are incomplete in the source data and may contain rule-based completion.
- Multiple inventors or applicants may be represented as delimited values. Users must normalize these fields before person-level or country-level counting.
- A patent office identifies the filing authority, not the inventor's country.
- Recent priority years are subject to publication and database-update lags.
- PATSTAT coverage and field definitions vary by office, jurisdiction, and year.
- CPC is less complete than IPC in the integrated master.

## Privacy and Responsible Use

Patent records may contain names and address-related fields. Do not use this dataset to profile, contact, rank, or make consequential decisions about individuals. Apply data-minimization and applicable privacy rules. Users should independently review redistribution rights, personal-data handling, and institutional requirements before copying or republishing the dataset.

## License and Access

No open-data license is asserted for the underlying PATSTAT-derived records. Access to and reuse of PATSTAT content remain subject to the applicable EPO/PATSTAT terms and the uploader's institutional permissions. Public repository access does not transfer ownership, grant additional reuse rights, or waive third-party rights.

## Reproducibility and Verification

The repository includes processing code, integration reports, file-level SHA-256 checksums, and a technical report. The packaging script creates Zstandard-compressed Parquet files and a manifest containing row counts, byte sizes, and checksums. Verify the manifest after downloading.

## Citation

If permitted to use this dataset, cite the repository version or commit hash, the accompanying technical report, and EPO PATSTAT Online 2026 Spring. Add project authorship and institutional citation details before external distribution.