Open data · 2,516,198 places · 524 cities · 106 countries
The open place corpus
Every place in our catalogue, as one download: 2,516,198 places in 524 cities across 106 countries, as Parquet with a real geometry, ranked and attributed. One file for the world, or one shard per city if you only need Paris. It is the same catalogue the ranked lists are built from — free to download, because a corpus an agent cannot cite is a corpus that does not exist.
Download
| File | Size | Rows | sha256 |
|---|---|---|---|
| manifest.json | 388 KB | — | ce2ee4759071… |
| world/places.parquet | 292.0 MB | 2,516,198 | 5f7d39b8307e… |
| sample.json | 18 KB | 20 | f937be692ace… |
| world/place_source.parquet | 154.8 MB | 5,573,762 | 80aeed75d298… |
| world/contributions.parquet | 194 B | 0 | ae522e0f3d74… |
Every shard is GeoParquet 1.0: a WKB geometry column (WGS84) beside
lon/lat, zstd-compressed. The object key contains the file's own sha256,
so a rebuilt dump is a new URL — never different bytes behind a URL your cache already holds.
curl -s https://data.555.team/v1/manifest.json | jq '.totals'
curl -O https://data.555.team/v1/world/places-5f7d39b8307e.parquet
python -c 'import duckdb; print(duckdb.sql("select name, city_id, rank_score from 'places-5f7d39b8307e.parquet' order by rank_score desc limit 5").fetchall())'
One city, not the world
Each city is its own shard at https://data.555.team/v1/city/<geonameid>-<slug>/places-<sha256[0:12]>.parquet — the median city is
66 KB, the largest (tokyo-jp) 11.3 MB. The manifest lists all 524; the twelve biggest:
| City | Places | Size |
|---|---|---|
| Tokyo tokyo-jp | 111,373 | 11.3 MB |
| New York City new-york-city-us | 100,890 | 8.9 MB |
| Los Angeles los-angeles-us | 85,020 | 7.6 MB |
| São Paulo sao-paulo-br | 69,568 | 6.8 MB |
| Chicago chicago-us | 67,842 | 6.1 MB |
| Houston houston-us | 56,931 | 5.2 MB |
| London london-gb | 51,599 | 5.2 MB |
| Dallas dallas-us | 55,004 | 5.0 MB |
| Osaka osaka-jp | 47,404 | 4.8 MB |
| Paris paris-fr | 42,884 | 4.3 MB |
| Taipei taipei-tw | 37,881 | 3.7 MB |
| Bangkok bangkok-th | 31,814 | 3.3 MB |
What licence applies to what
The corpus is a compilation of 10 open sources plus our own derived fields,
so the licence is per field group, not per file. Each source's own licence is also
recorded per place in place_source.parquet, so a row can be checked against the provider
it came from.
| Field group | Columns | Licence | Attribution |
|---|---|---|---|
place.core | place_id name category basic_category category_path confidence address locality country_code website phone brand_name brand_qid | per provider — see `sources` | The providers' own fields. Each provider's licence is recorded per place in place_source, and the union is in sources. |
place.geo | lon lat geometry | per provider — see `sources` | WGS84. geometry is a WKB POINT with the GeoParquet 1.0 geo metadata, so the file opens as a spatial layer. |
place.city | city_id | CC-BY-4.0 | City assignment: GeoNames cities15000 (CC BY 4.0) |
place.derived | first_seen last_attested_at seed_release rank_score | ODbL-1.0 | maps.555.team — derived catalogue fields |
place.provenance | place_id source provider dataset licence external_id update_time confidence | per provider — see `sources` | place_source: one row per (place, provider, external id). The citation layer — a place can be checked against the provider it came from. |
contribution | place_id body attribution licence created_at | ODbL-1.0 | maps.555.team members — contributions published under ODbL-1.0 |
| Source | Licence | Rows |
|---|---|---|
| AllThePlaces | CC0-1.0 | 32,496 |
| BrightQuery | CDLA-Permissive-2.0 | 177,174 |
| DAC | CDLA-Permissive-2.0 | 11,083 |
| Foursquare Open Source Places | Apache-2.0 | 59,904 |
| Krick | CDLA-Permissive-2.0 | 8 |
| Meta | CDLA-Permissive-2.0 | 2,002,412 |
| Microsoft | CDLA-Permissive-2.0 | 221,086 |
| Overture Maps Foundation | CDLA-Permissive-2.0 | 3,057,564 |
| PinMeTo | CDLA-Permissive-2.0 | 11,758 |
| RenderSEO | CDLA-Permissive-2.0 | 277 |
The ODbL notice
Contains information from the maps.555.team place corpus, which is made available here under the Open Database License (ODbL) v1.0. Any rights in individual contents of the database are licensed under the Database Contents License (DbCL) v1.0.
The contribution group is published under ODbL-1.0 by the author's grant, and our derived fields are offered under the same licence. ODbL §4.6 is why the whole database is downloadable here, free of charge, while the service on top of it may be charged for.
The complete database is offered free of charge over the internet at https://data.555.team/v1/manifest.json
How to cite it
maps.555.team (2026). Open place corpus, version v1 (build 8d2acae0423f). https://maps.555.team/data
@misc{maps555corpus, title = {maps.555.team open place corpus}, author = {{maps.555.team}}, year = {2026}, note = {Version v1, build 8d2acae0423f, ODbL-1.0}, url = {https://maps.555.team/data}}
How fresh it is
On demand today — a build is cut from the live catalogue when the data changes, and a nightly rebuild is the next step. Every build is content-addressed, so a schedule changes nothing about the URLs: `build.built_at` is the instant of the run and `build.id` is a hash of the shard bytes, which means two builds over unchanged data are the same build. This build is 8d2acae0423f, produced
2026-10-21T19:55:40Z in 65.3s from the live catalogue. The exact command:
uv run --quiet --with duckdb --with 'psycopg[binary]' python pipeline/seed/dump.py --dsn "$CATALOG_DSN"
PostgreSQL `maps_catalog` (CNPG maps/maps-catalog): `public.place`, `public.place_source`, `app.contribution` WHERE status = 'published' · DuckDB (zstd Parquet, spatial extension for WKB geometry)
What is not in it
No ratings, no review counts, no prices and no opening hours: the catalogue has no licensable source
for them, so there are none here rather than a column a reader fills in. No private data either — a
member's lists, their favourites and their list notes are never published, and a contribution appears
only once it is published, attributed by a public name or a neutral member-… handle, never
by an email address.
What the ranking is
rank_score is our own curation-quality score: how well a place is attested by
independent providers, how complete its record is, and Overture's existence confidence. It is not
popularity and nobody pays for placement. The score, its inputs and its anchor validation are in the
data; the model's weights are not (ODbL §2.3(a) leaves the program outside the licence).