Worldwide OSM polygons with textual descriptions from the “description” tags

Dataset briefing · snapshot 2026-08

906,631 deduplicated polygon rows, each carrying the source tags, geometry, area, provenance, and exact description text.

Open dataset · GeoParquet 1.1 · ODbL-derived OpenStreetMap data

Sources: generated stats.json; docs/dataset-contract.md; dataset card snapshot

Which polygons do we keep?

Keep a polygon if it has at least one non empty description tag.

Included

Tagged closed ways that OSM classifies as areas
type=multipolygon and type=boundary relations that assemble
description=* or description:<suffix>=*
area=yes remains authoritative

Excluded

Nodes and open ways
area=no
Empty or whitespace-only descriptions
Failed relation assemblies and non-polygon output

Sources: docs/dataset-contract.md; config/osmium-export.json; generated stats.json

The snapshot is large in input, selective in output

83.0 GB

source PBF bytes

869.9M

emitted OSM features

906.6k

published rows

386 regional PBF extracts become 386 Parquet files totaling 712.4 MiB. The final rows are unique by (osm_type, osm_id); 38,851 duplicate candidates were removed during global deduplication.

Source: generated stats.json; values rounded for presentation, exact values remain in stats.json

Every row keeps the evidence behind the label

Text and provenance

Base and localized names
Base and localized descriptions
Complete original tags map
OSM version, changeset, timestamp, source PBF, and URL

Geometry and measurement

Full Polygon or MultiPolygon WKB
OGC:CRS84 longitude/latitude semantics
WGS84 geodesic area_m2
Bounding box covering the complete geometry

Sources: docs/dataset-contract.md; generated stats.json

Descriptions are predominantly base text, with a multilingual tail

Base descriptions

887,077

values · 5.11M words · median 4 words

Localized descriptions

32,049

values · 214k words · median 3 words

The most common exact localized suffixes are de, en, it, fr, and ru. Suffixes are preserved verbatim; they are not asserted to be valid language codes.

Source: generated stats.json; suffix interpretation from docs/dataset-contract.md

The global distribution is dense but uneven

H3 density of description-tagged polygons

7,381 H3 resolution-3 cells contain rows.

Source: assets/description_polygon_density.png; docs/dataset-contract.md; generated stats.json

Most polygons are small, but the tail reaches country scale

Area distribution of description-tagged polygons

Median area: 501 m² · middle 50%: 85–10,391 m² · range: 0.000062–3.48×10¹² m².

Source: assets/area_distribution.png; generated stats.json; area_m2 is WGS84 geodesic area

The timestamps span the history of modern OSM mapping

2007 → 2026

minimum to maximum OSM object timestamp (UTC)

The timestamp is object provenance, not a dataset publication date. Text and geometry reflect the source extracts at their recorded OSM timestamps.

Source: generated stats.json; docs/dataset-contract.md

Interpretation needs three guardrails

Language — localized suffixes are exact OSM keys, not validated language codes.

Text quality — descriptions are community-authored and vary in language, formatting, completeness, and quality.

Spatial meaning — the H3 map counts polygon centroids and the extracts are regional snapshots.

Sources: dataset-card-template.md; docs/dataset-contract.md

Resources

Access and reuse

Dataset: huggingface.co/datasets/NoeFlandre/osm-polygon-description-tag

Metrics: Trackio snapshot dashboard

Github: github.com/NoeFlandre/osm-polygon-description-tag

Sources: generated README.md; docs/index.md; docs/dataset-contract.md