Dataset Validator

mlcast-dataset-validator is the validation tool for datasets contributed to the MLCast Intake catalog (mlcast-datasets). It ensures a dataset meets the technical requirements for inclusion in the MLCast data collection before submission.

Full specification docs: https://mlcast-community.github.io/mlcast-dataset-validator/

Why it exists¶

During the MLCast community meeting, multiple entities offered to contribute datasets. To streamline contribution and ensure data quality, the validator lets providers verify their Zarr archives are compliant before submitting. It addresses two needs:

1. Specification compliance — validates datasets against the formal MLCast Zarr format specification v1.0 (RFC 2119 keywords).

2. Tool compatibility — tests that datasets work correctly with common geospatial tools (xarray, GDAL, cartopy).

What it validates¶

Currently covers radar precipitation source datasets.

Minimum requirements for acceptance

• 2D radar composite at 1 km resolution or finer

• at least 256×256 pixel valid sensing area

• minimum 3 years of temporal coverage

• consistent spatial domain across all timesteps

• data variable in mm (depth), mm/h (rate), or dBZ (reflectivity)

Technical requirements

• GeoZarr format (Zarr v2/v3 with proper georeferencing)

• CF-compliant coordinate and variable names

• correct dimension ordering (time × H × W)

• proper chunking (1 chunk per timestep)

• ZSTD compression (recommended)

• NaN values for missing / out-of-range data

• license metadata (CC-BY, CC-BY-SA, OGL, etc.)

Tool compatibility

• xarray can load and slice the data correctly

• GDAL can interpret the georeferencing (WKT parsing)

• cartopy can create CRS objects and transform coordinates

• cross-tool CRS consistency checks

How it is organized¶

• Spec modules by data stage / product — mlcast_dataset_validator/specs/<data_stage>/<product>.py (e.g. specs/source_data/radar_precipitation.py). specs/training_data/ is prepared for future ML training datasets.

• Spec sections mirror an xr.Dataset — each module follows the dataset layout (coordinates, variables, global attrs, tool compatibility).

• Inline spec text drives each requirement — human-readable RFC 2119 spec text sits next to the function calls that enforce it.

• Reusable checks — under mlcast_dataset_validator/checks/<section>/<aspect>.py.

mlcast_dataset_validator/
├── specs/
│   ├── source_data/
│   │   └── radar_precipitation.py
│   ├── training_data/          # no specs yet
│   └── cli.py
└── checks/
    ├── coords/                 # names, spatial, temporal, variable_timestep
    ├── data_vars/
    ├── global_attributes/
    └── tool_compatibility/

Usage¶

From the command line¶

Easiest path: run the PyPI release directly with uv:

uvx --from mlcast-dataset-validator mlcast.validate_dataset <data_stage> <product> <dataset-path>

Validate a local Zarr dataset:

uvx --from mlcast-dataset-validator mlcast.validate_dataset \
    source_data radar_precipitation /path/to/radar_precip_source.zarr

Validate a remote Zarr in an S3 bucket at a custom endpoint (here the Radklim Zarr already in the intake catalog):

uvx --from mlcast-dataset-validator mlcast.validate_dataset \
    source_data radar_precipitation \
    s3://mlcast-source-datasets/radklim/v0.1.1/5_minutes.zarr/ \
    --s3-endpoint-url https://object-store.os-api.cci2.ecmwf.int --s3-anon

Or clone and run directly:

git clone https://github.com/mlcast-community/mlcast-dataset-validator
cd mlcast-dataset-validator
pip install -e .
mlcast.validate_dataset source_data radar_precipitation /path/to/zarr/file.zarr

From Python¶

Import the relevant spec and call it with an xr.Dataset. This is how the validator runs in the CI of the mlcast-datasets repo, validating datasets on every PR and main-branch commit.

import xarray as xr
from mlcast_dataset_validator.specs.source_data import radar_precipitation

storage_options = {
    "endpoint_url": "https://object-store.os-api.cci2.ecmwf.int",
    "anon": True,
}

ds = xr.open_zarr(
    "s3://mlcast-source-datasets/radklim/v0.1.1/5_minutes.zarr/",
    storage_options=storage_options,
)
# Preserve storage options so zarr_format checks can inspect remote-store metadata.
ds.encoding.setdefault("storage_options", storage_options)

report, _ = radar_precipitation.validate_dataset(ds)
report.console_print()