Datasets:

---

Giezi

/

MicroWeather-Bench

Like 2

This PR adds a dataset card (README.md) so that the repo page explains what MicroWeather-Bench contains and how to load it. Everything in the card was taken from the stores' own metadata (dimensions, attributes, coordinate ranges, chunking) and from the station files, and every code example was run against the repo. The YAML header keeps configs: [] and viewer: false.

Please treat it as a draft to edit. A few things for the maintainers to decide or fix:

• License. No license key was added because none has been chosen for the compiled dataset. The card lists each upstream source's terms with a link instead, checked against the sources' own pages.
• MADIS redistribution policy. MADIS allows redistribution only of category-3 ("Public - Full distribution") provider data, asks that NOAA's MADIS team (NCEP.List.MADIS_Support@noaa.gov) be notified before data is placed on public web pages, and asks that files keep the data-provider information. The card states the policy neutrally. It would be worth confirming that all stations come from category-3 providers, sending the notification if it has not been sent, and considering a per-station provider field in stations/locations.geojson.
• Split grouping. stations/splits.json is a station-level split: 1277 of the 4741 occupied ERA5 0.25-degree cells contain stations from more than one group, and 301 of the 528 cells with a test station also hold a non-test station. If the benchmark is meant to score held-out ERA5 cells, the split needs to be grouped by cell. The card describes the split neutrally.
• Metric. The card says the repo fixes the data and the split but not an error metric. If you want to fix one (and the aggregation), the "Task and protocol" section is the place.
• Optional header keys. task_categories was left out; tabular-regression is the closest Hub category if you want one.
• Store metadata nits (not mentioned in the card): terrain/COPERNICUS_DEM_GLO30_DEM-CONUS.zarr has the title "... - Slope" although it holds the DEM; terrain/JAXA_ALOS_AW3D30_V3_2-CONUS.zarr has institution: ESA Copernicus (the source is JAXA); the TESSERA store's history says "TESSSERA"; the MADIS store's spatial_coverage, temporal_coverage and institution are one-element lists while every other store uses plain strings; the two TerraMind stores share the same title; the grid_mapping: spatial_ref attribute on the patch stores points to a variable that does not exist.