taylor-geospatial/CoordBench

CoordBench

A unified benchmark suite for evaluating location encoders such as SatCLIP, GeoCLIP, Climplicit, and MIND. The dataset contains 40 normalized source tables from 13 source families. The paper's evaluation suite uses 52 datasets and 78 prediction targets drawn from this mirror. The source files previously lived across GitHub, figshare, GCS, Socrata, Zenodo, and Google Drive.

Intended use

Use the normalized tables to compare coordinate-to-embedding models. The HF config is defined by the original source table, not by individual targets. The paper evaluates 52 datasets and 78 targets with random five-fold cross-validation and latitude--longitude regional holdouts.

The normalization pipeline preserves source columns where possible, but it also joins coordinate columns, creates procedural samples for polygon and raster sources, and applies source-specific parsing. Read each config's provenance notes for more information.

The benchmark does not provide a universal license. Please refer to the per-config license table and the original source terms before redistribution or commercial use.

Load one table

import pyarrow.parquet as pq
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    "XXXX/CoordBench",
    "data/california_housing/data.parquet",
    repo_type="dataset",
    token=False,
)
table = pq.read_table(path, columns=["lon", "lat", "MedHouseVal"])
print(table.num_rows, table.column_names)

The remaining configs use the same data/<config>/data.parquet layout.

Schema

Every config's data.parquet carries these columns up front:

...followed by the source columns retained by the normalization pipeline. Nodata values (-999, -1, ocean index 0) and untransformed values (no log1p) are preserved where the source table is copied. The config notes identify joins, procedural samples, and source-specific parsing.

Note there's one config per original source file/table, not per label task e.g. sustainbench's 6 DHS indices and cdc_places's 12 health measures are each one config since they're one shared source table.

A raw/ folder alongside data/ contains the original unmodified copies of the original datasets files.

License

No single license applies to the whole repo. Most configs are CC-BY-4.0/CC-BY-SA-4.0/public domain however some are unique -- see the status column below:

Comparison with MIND

Note that the bt-* and country and ecoregions datasets were added at a later date to CoordBench so to reproduce MIND results from the paper exclude these from your evals.