Hugging Face's logo

Datasets:
phatcher
/
companyblocker

Modalities:
Text
Formats:
parquet
Languages:
English
French
German
Size:
10M - 100M
Tags:
entity-resolution
record-linkage
blocking
company-names
DOI:
Libraries:
Datasets
Dask
Polars
License:
Dataset card Data Studio Files
xet
 Community
1
companyblocker / README.md
phatcher's picture
phatcher
Data card describing the dataset:
4ecc096
preview code
|
Raw History Blame Contribute Delete
10.7 kB
metadata 
license: mit
language:
  - en
  - fr
  - de
tags:
  - entity-resolution
  - record-linkage
  - blocking
  - company-names
size_categories:
  - 10M<n<100M
configs:
  - config_name: companies
    description: One row per company, in the shared canonical schema, split by source.
    data_files:
      - split: fr
        path: data/canonical/fr/primary/**/*.parquet
      - split: gb
        path: data/canonical/gb/primary/**/*.parquet
      - split: de
        path: data/canonical/offeneregister/primary/**/*.parquet
      - split: ie
        path: data/canonical/ie/primary/**/*.parquet
      - split: gleif
        path: data/canonical/gleif/primary/**/*.parquet
      - split: wikidata
        path: data/canonical/wikidata/primary/**/*.parquet
  - config_name: names
    description: Other names each company is known by, split by source.
    data_files:
      - split: fr
        path: data/canonical/fr/names/*.parquet
      - split: gb
        path: data/canonical/gb/names/*.parquet
      - split: de
        path: data/canonical/offeneregister/names/*.parquet
      - split: wikidata
        path: data/canonical/wikidata/names/*.parquet
      - split: gleif
        path: data/canonical/gleif/names/*.parquet
  - config_name: matched
    description: >-
      GLEIF rows carrying a link into a national register, split by
      jurisdiction.
    data_files:
      - split: de
        path: data/matched/gleif/primary/jurisdiction_code=de/*.parquet
      - split: fr
        path: data/matched/gleif/primary/jurisdiction_code=fr/*.parquet
      - split: gb
        path: data/matched/gleif/primary/jurisdiction_code=gb/*.parquet
      - split: ie
        path: data/matched/gleif/primary/jurisdiction_code=ie/*.parquet
  - config_name: samples
    description: About 16k rows per source, to look at before downloading a whole one.
    data_files:
      - split: fr
        path: data/canonical/fr/sample.parquet
      - split: gb
        path: data/canonical/gb/sample.parquet
      - split: de
        path: data/canonical/offeneregister/sample.parquet
      - split: ie
        path: data/canonical/ie/sample.parquet
      - split: gleif
        path: data/canonical/gleif/sample.parquet
      - split: wikidata
        path: data/canonical/wikidata/sample.parquet
  - config_name: duplicates
    description: >-
      Rows dropped as duplicates, kept for inspection; two columns more than a
      company row.
    data_files:
      - split: gleif
        path: data/canonical/gleif/gleif-duplicates-*.parquet
      - split: ie
        path: data/canonical/ie/ie-duplicates-*.parquet

Company registers in one schema, with name variants and cross-register links

Six public company sources mapped into one column set, with the name variants each register publishes and the GLEIF rows that carry a link into a national register. Built for blocking and record linkage.

It holds:

  • 884,070 Wikidata companies, extracted from the 145 GB entity dump by the P279 subclass closure beneath company.
  • Six registers on one column set, 28.7M companies in total.
  • 8.1M name variants, as each register publishes them.
  • 645,437 GLEIF rows carrying a link into a national register.

What is here

Source Rows Names Jurisdictions Snapshot
fr, INSEE SIRENE 12,930,798 1,397,729 fr 2026-06-01
gb, Companies House 5,698,275 645,313 gb 2026-06-01
offeneregister, German Handelsregister 5,305,727 1,346,089 de 2026-08-25
gleif, Global LEI index 3,076,432 3,638,577 308 2026-06-21
wikidata 884,070 1,054,902 73 2026-07-16
ie, CRO Ireland 817,761 — ie 2026-06-21

Each source is the publisher's own bulk file, taken once on the date given.

Layout 

data/canonical/<source>/primary/jurisdiction_code=<code>/part-*.parquet   one row per company
data/canonical/<source>/names/*.parquet                                   other names for those companies
data/canonical/<source>/sample.{parquet,jsonl,csv}                        ~16k rows, the same rows in three formats
data/canonical/<source>/*-duplicates-*.parquet                            rows dropped as duplicates, kept for inspection
data/matched/gleif/primary/jurisdiction_code=<code>/part-*.parquet        GLEIF rows carrying a national-register link
data/prepare/wikidata/wikidata-companies.jsonl                            the companies drawn out of the Wikidata dump

In the CSV the two list columns, alternative_names and previous_names, are joined with |; the JSON Lines copy keeps them as lists.

The Wikidata extract

wikidata-companies.jsonl is the raw extract: 884,070 entities in 612 MB, one JSON object per entity as the dump holds it, before any mapping. It was taken from the full 145 GB entity dump in a single pass.

An entity is kept when its P31 (instance of) or P1454 (legal form) names a Q-code in the P279 subclass closure beneath company, 49,176 codes. Everything under data/canonical/wikidata/ is built from this file, which also carries the fields the canonical schema does not keep.

primary is partitioned by jurisdiction_code, so one country can be read without the rest:

import polars as pl

gb = pl.read_parquet("data/canonical/gb/primary/jurisdiction_code=gb/*.parquet")
names = pl.read_parquet("data/canonical/gb/names/*.parquet")

Columns

A company (primary, 21 columns)

Column names follow the OpenCorporates company object where one fits.

Column Description
system_uri This row's identity, <source>://<id>, e.g. gb://03909510
jurisdiction_code Where the company is registered
company_number The register's own identifier for it
name Its legal name
alternative_names, previous_names Other names, where the register gives them inline
company_type Legal form, as the register words it
current_status, inactive Whether it is live, as the register words it
incorporation_date, dissolution_date ISO dates where parsed
branch, branch_status Whether the row is a branch rather than a company
registered_address_in_full The address as one string, where published
registry_url Its page in the register
lei Legal Entity Identifier, where one resolves
match_uri The same company in another source, where one is known
opencorporates_url, vat, tax_id Reserved, always null
metadata_generated_utc When this row was written here

A name (names, 9 columns)

One row per name a company is also known by.

Column Description
system_uri This name's own identity, name://<source>/<id>/<hash>
source_uri The company it belongs to
name The name itself
name_type What kind: previous, alias, official, short, acronym, ...
source_type What the register called that kind, before it was mapped
language_code, jurisdiction_code, id, derivation_note Where it came from

A cross-register link (data/matched/gleif)

The same 21 columns, holding the 645,437 GLEIF rows whose match_uri points at a company in gb, ie, fr or de.

How it was built

Each source's bulk file was downloaded, split into shards, and mapped into the columns above, with per-source rules for the parts that differ: how a company number is spelled, which field is the legal name, which value means dissolved. Duplicate rows were removed and kept separately rather than dropped silently. Names were gathered from whichever field each register publishes them in and given one shared vocabulary of name types.

Names are as the registers publish them: not normalised, lowercased or stripped of their legal form.

Known gaps

  • ie has no names file. Ireland's bulk download publishes one name per company, so there are no variants to collect.
  • wikidata rows with no jurisdiction sit in a __HIVE_DEFAULT_PARTITION__ partition.
  • Every source is one snapshot, taken on the date in the table. Companies registered, renamed or dissolved since are not here.
  • match_uri is only as complete as GLEIF is. A company with no LEI has no link, so absence of a link is not evidence that no match exists.

Licensing information

Each source keeps its own terms. What each publisher states, in its own words, with the page it says it on. Those pages are the terms; this table points at them.

Source Licence, as the publisher states it Stated on
fr Licence Ouverte / Open Licence 2.0 data.gouv.fr
gb "All content is available under the Open Government Licence v3.0, except where otherwise stated" Companies House data products
offeneregister "Die Daten können unter der Creative Commons Attribution License 4.0 mit Attribution an OpenCorporates weiterverwendet werden" offeneregister.de
gleif "The data available through the Access Service are provided under the CC0 licence" LEI data terms of use
wikidata "All structured data from the main, Property, Lexeme, and EntitySchema namespace is available under the Creative Commons CC0 License" Wikidata:Database download
ie Creative Commons Attribution 4.0 CRO open data

Two publishers ask for attribution in particular words:

  • German data: attribution to OpenCorporates.
  • Irish data: "Contains Irish Public Sector Data licensed under a Creative Commons Attribution 4.0 International (CC BY 4.0) licence".

The license field at the top of this card is the repository's own metadata field and restates none of the above. Statements read from those pages on 2026-09-23; publishers change their terms, so the pages are the current source.

What is in the rows

These are public company registers, so the rows carry what registers carry: company names, registered addresses, and legal forms.

French sole traders are not among them. SIRENE's legal category 1000, Entrepreneur individuel, is a natural person rather than a company, and those rows are dropped when the French source is mapped into this schema: no row under fr carries that category.