MLCast is split across several repositories under
github
Datasets (mlcast-datasets) — an Intake catalog of curated, mlcast-compliant source radar datasets, with per-provider converters and a sampler that turns source data into training-ready indices.
Validator (mlcast
-dataset -validator) — checks that a contributed dataset meets the MLCast Zarr specification and works with common geospatial tools before it joins the catalog. Machine learning (mlcast) — the main Python package: ConvGRU ensemble nowcasting models and a configuration-driven CLI / Python API for training.
The typical flow: a dataset is converted and validated, registered in the catalog, sampled into training indices, then fed to the mlcast package for training.
Quick start¶
Get data — open the Intake catalog of curated radar datasets. Install from GitHub for the latest version, or
pip install mlcast-datasetsfor a stable release. See Datasets.Validate (contributors only) — if you are adding a new dataset, check it against the MLCast Zarr spec first. See Validator.
Train a model — fork and clone mlcast, install with
uv sync, then runmlcast train. See Usage.