ne-admin0-10m

Natural Earth's boundaries at 1:10m, frozen at version 5.1.1, with the archives they were built from and the checksums that say so.

from datasets import load_dataset
countries = load_dataset("yuiseki/ne-admin0-10m", "v5.1.1.admin0")   # 258
states    = load_dataset("yuiseki/ne-admin0-10m", "v5.1.1.admin1")   # 4,596
counties  = load_dataset("yuiseki/ne-admin0-10m", "v5.1.1.admin2")   # 3,224, US only

subset	layer	features	attributes
v5.1.1.admin0	ne_10m_admin_0_countries	258	168
v5.1.1.admin1	ne_10m_admin_1_states_provinces	4,596	121
v5.1.1.admin2	ne_10m_admin_2_counties	3,224	61

The hierarchy closes

Every one of the 4,596 states names a country in adm0_a3, and every one of those is in the admin-0 layer. No orphans. That is what makes transitivity usable on real data rather than on a worked example: a state is within a country, a country is within nothing else here, and both steps exist.

Japan has its 47 prefectures. Kagoshima is JP-46 and Q15701, the same item a gazetteer would use.

The third rung is one country

admin-2 is the United States and nothing else. 3,224 counties, and Natural Earth publishes no equivalent for the other 257 countries at this scale. A hierarchy built on it works for one country and silently has nothing to say about anywhere else, which is worth knowing before building one.

Not all of them are counties either: 3,008 are, plus 77 municipios of Puerto Rico, 64 parishes of Louisiana, 41 independent cities, 13 boroughs and 11 census areas of Alaska. TYPE_EN says which.

3,212 of the 3,224 carry NAME_JA, so Whatcom County is ワットコム郡, and all 3,224 carry WIKIDATAID.

The join that looks obvious and is not

Both layers have a column called iso_3166_2 and they do not agree. admin-1 writes US-WA and admin-2 writes US-53: the postal abbreviation against the FIPS number. Joining on the shared column name matches nothing at all, and matches it quietly.

REGION is the column that works. It holds WA, which is the second half of admin-1's US-WA.

postal = {r["iso_3166_2"].split("-", 1)[1]: r
          for r in states if r["adm0_a3"] == "USA"}
county_to_state = [(c, postal[c["REGION"]]) for c in counties
                   if c["REGION"] in postal]

That joins 3,143 of 3,224. The 81 left are Puerto Rico's 78 and the US Virgin Islands' 3, which Natural Earth carries as admin-0 entries rather than as states of the United States, so they resolve one rung up instead.

Why a frozen copy exists

Natural Earth publishes one URL per layer, and that URL always serves the current release:

https://naciscdn.org/naturalearth/10m/cultural/ne_10m_admin_0_countries.zip
https://naciscdn.org/naturalearth/10m/cultural/ne_10m_admin_1_states_provinces.zip

There is no way to ask either for 5.1.1. A result computed against "Natural Earth admin-0 10m" is therefore not reproducible by naming the source, which is the kind of gap that goes unnoticed until a boundary moves.

What freezes it is the checksum, not the copy. Each archive declares its own version in a VERSION.txt inside it, which the builder reads and checks rather than trusts, and each is pinned by size, md5 and sha256:

layer	bytes	md5
admin-0	4,930,492	0bc33f3e6f9ab457917430e57f6122a9
admin-1	14,909,524	63b322fddc41d4c6877eafe4e763c897
admin-2	1,919,144	81fc622733e1e89d48bdb6b292d21bec

The same bytes are reachable a second way, from the v5.1.1 tag of the upstream build repository at commit 9380cca8.

What is here

v5.1.1/ne_10m_admin_0_countries.zip                 the publisher's archives,
v5.1.1/ne_10m_admin_1_states_provinces.zip          untouched
v5.1.1/*.zip.md5
v5.1.1/VERSION.txt                                   5.1.1
v5.1.1/parquet/ne_10m_admin_0_countries.parquet
v5.1.1/parquet/ne_10m_admin_1_states_provinces.parquet
v5.1.1/parquet/checksums.md5
v5.1.1/manifest.json                                 counts and digests
v5.1.1/provenance.yaml                               where it came from

The Parquet files are the same records: every attribute column under its own name, nothing renamed and nothing dropped, with the geometry beside them as WKB in a column called geometry.

Wikidata, and a trap

WIKIDATAID is set on all 258 countries, including the cases where a boundary is argued over: Taiwan Q865, Palestine Q23792, Kosovo Q1246.

On admin-1 the column is wikidataid, in lower case. Natural Earth spells it differently between the two layers and this repository does not normalise it, because a query written for one then fails on the other instead of silently returning nothing.

It is also not complete: 4,332 of 4,596 states carry one, so a join on it drops 264 rows. adm0_a3 and iso_3166_2 are set on all of them.

The four provenance columns

column	admin-0	admin-1
geometrySource	Natural Earth	Natural Earth
datasetVersion	5.1.1	5.1.1
boundaryView	default/de-facto	default/de-facto
adminLevel	admin-0	admin-1

Each is constant within a layer, and each stops being constant the moment something is added beside these rows: a second source, a second version, or one of the 33 national points of view Natural Earth also ships. A table that will hold those has to carry the columns from the start, because added later the older rows have already lost the answer.

boundaryView is the one that matters most. Natural Earth draws de facto boundaries by default, according to who controls the territory. The _chn, _jpn, _ind, _pak, _isr, _pse and other variants are not here; the column is what would let them be added beside these rows rather than instead of them.

Licence

Public domain. Natural Earth asks no permission and requires no attribution, though it appreciates credit:

Made with Natural Earth. Free vector and raster map data @ naturalearthdata.com

This is unusual and worth knowing. Unlike anything derived from OpenStreetMap, which is ODbL and whose share-alike reaches a derived database, this can be mixed with other sources without changing their terms.

See v5.1.1/provenance.yaml for every checksum and v5.1.1/LICENSE for the full statement.