OSMGraphCLIP-MS

Training dataset for the MS (multiscale) variant of OSMGraphCLIP: "OSMGraphCLIP: Learning Global Location Representations from OpenStreetMap Graphs" (arXiv:2606.08046).

It contains ~200k globally-diverse (lat, lon) locations, and for each location:

• a heterogeneous OSM graph (points, lines, polygons — roads, buildings, land use, POIs — with SBERT node features), and
• multiscale concentric-ring "band" features summarizing OSM content in rings around the location at multiple radii.

These pairs (graph + coordinate, with the band features as auxiliary multiscale signal) are what the graph encoder (OSMHeteroGAT) and location encoder (LocationEncoder) are contrastively aligned on. This dataset was used to train:

Model	HuggingFace
OSMGraphCLIP-MS-L40	d-michail/OSMGraphCLIP-MS-L40
OSMGraphCLIP-MS-L10	d-michail/OSMGraphCLIP-MS-L10

Code to build datasets in this format, and to train on them, is in the osmgraphclip repo (create_dataset.py, create_multiscale_dataset.py, create_graphs.py, train.py).

Subsets

The dataset is split into two location sets, matching the two location CSVs in the code repo:

Subset	Locations	Source	Samples with no OSM data
satclip_ms/	100,000	data/satclip_locations.csv — primary location set	1,076
h3_ms/	99,260	data/h3_locations.csv — globally-diverse H3-sampled locations	645

Each subset is independent and has the same internal layout.

Repo layout

{h3_ms,satclip_ms}/
├── metadata.json           # generation config for this subset (see below)
├── dataset.db               # SQLite index (see below)
├── graphs.zip-parts/        # sharded graphs/ folder (~1 GiB zip shards)
│   ├── graphs-000.zip
│   ├── graphs-001.zip
│   └── ...
└── bands.zip-parts/         # sharded bands/ folder (~1 GiB zip shards)
    ├── bands-000.zip
    └── ...

The raw graphs/ and bands/ folders are not stored directly in the repo (hundreds of thousands of small files each — unfriendly to git/HF). They are shipped as a sequence of zip shards instead. Each shard stores entries as <subfolder>/<filename>, so unzipping every shard for a given subset/subfolder combo into that subset's root directory reconstructs the original folder exactly, with no overlap between shards.

Reconstructing graphs/ and bands/

Use unpack_shards.py from this repo:

python3 unpack_shards.py h3_ms satclip_ms

This walks the given root(s), finds every *.zip-parts/ directory, and unzips all shards inside it into the parent directory — reconstructing h3_ms/graphs/, h3_ms/bands/, satclip_ms/graphs/, satclip_ms/bands/. It's safe to re-run.

graphs/ contents

For each location, identified by an integer id <id>:

• osm_<id>_graph.pkl — a pickled torch_geometric.data.HeteroData object: the heterogeneous OSM graph with node types polygon (392-dim features), line (390-dim), point (386-dim), and all 9 directed edge-type combinations between them (edge_index + edge_attr). Node features are SBERT (all-MiniLM-L6-v2, 384-dim) embeddings of OSM tags augmented with per-type geometric attributes. Built with the GeoLink-derived osm_to_graph.py pipeline (graph_method: geolink).
• osm_<id>_{point,linestring,multilinestring,polygon,multipolygon}.geojson.gz — the raw gzipped OSM GeoJSON geometries (with tags) that the graph for that location was built from. Not every geometry type is present for every location.
• osm_<id>.nodata — present instead of the geojson files when no OSM data was found in the location's bounding box. The corresponding graph is still written (an empty HeteroData, method = "zero" in dataset.db), so <id> always has a valid graph pickle — .nodata just flags "empty, not missing/corrupted".

bands/ contents

• osm_<id>_bands.npz — multiscale concentric-ring band features at radii [2000, 10000, 20000] meters (see band_radii_m in metadata.json), keyed by:
  • band_radii — the 3 radii, in meters
  • spatial_features (3, 47) + spatial_feature_names — per-band aggregate spatial statistics
  • subbin_spatial (3, 2, 16) + subbin_feature_names — per-band, per-subbin (inner/outer half of the ring) spatial statistics
  • sector_spatial (3, 4, 11) + sector_feature_names — per-band, per-sector (quadrant) spatial statistics
  • global_embeddings (3, 384), subbin_embeddings (3, 2, 384), sector_embeddings (3, 4, 384) — SBERT embeddings of OSM tags aggregated at the whole-band / subbin / sector level

These give the graph encoder multiscale context beyond the single bounding box used for the main graph.

dataset.db

A SQLite database indexing every location by id, with three tables (same ids line up across tables and against the osm_<id>_* filenames):

• downloads: lat, lon, bbox_size, geojson_prefix (the osm_<id> prefix), timestamp — one row per raw OSM download.
• graphs: lat, lon, bbox_size, graph_pickle (filename), method (geolink or zero), timestamp — one row per built graph.
• band_features: lat, lon, bands_path (filename), band_radii, timestamp — one row per band-feature file.

metadata.json

Generation config shared by every sample in the subset:

{
  "bbox_size_m": 1000,
  "band_radii_m": [2000.0, 10000.0, 20000.0],
  "location_source": "data/h3_locations.csv",   // or data/satclip_locations.csv
  "tagw_path": "data/all_tags30_frequency1.json",
  "embedding_backend": "sbert",
  "graph_method": "geolink"
}

Citation

@misc{michail2026osmgraphcliplearninggloballocation,
      title={OSMGraphCLIP: Learning Global Location Representations from OpenStreetMap Graphs},
      author={Dimitrios Michail and Eleni Saka and Ioannis Giannopoulos and Ioannis Papoutsis},
      year={2026},
      eprint={2606.08046},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2606.08046},
}

License and acknowledgements

Contains data from OpenStreetMap, © OpenStreetMap contributors, available under the Open Database License (ODbL).