deepguess/severe-weather-env-all-models-v1

Severe Weather Environment All Models v1

This dataset is the convenience union of the HRRR, GFS, and RAP severe-environment packages. It is useful for broad experiments, model-system comparison, and quick loading of all current local severe-environment training rows.

This is a research dataset, not operational guidance. It is not an official NOAA, NWS, SPC, HRRR, RAP, or GFS product, and it should not be used as a stand-alone warning or forecast system.

Files

• data/all_models_training.parquet: union-schema table.
• data/preview_1000.csv: small preview.
• metadata/manifest.json: counts and source metadata.
• metadata/columns.json: column list.
• metadata/checksums.sha256: checksums.

Rows: 773,128
Columns: 993

Components

• hrrr_event_control: 370,160 rows
• gfs_point_forecast: 364,748 rows
• rap_tornado_environment: 38,220 rows

Every row has:

• dataset_component
• dataset_model_system
• dataset_row_role

Use those fields when training. HRRR, RAP, and GFS should not be treated as one identical feature distribution.

Suggested Uses

1. Train model-specific heads from one table by filtering dataset_model_system.
2. Train a shared representation with explicit model-system embeddings or one-hot model-system features.
3. Compare RAP and HRRR short-range environment skill.
4. Use GFS rows for lead-time calibration and HRRR/RAP rows for short-range environment discrimination.
5. Run feature availability audits across model systems.

QC And Construction

• The union table is built from already packaged local artifacts.
• No rows are hand-labeled during this packaging step.
• Missing values are expected where a feature exists for one model system but not another.
• RAP remains RAP-only; HRRR and GFS rows are not used to fill RAP failures.
• For strict scientific work, train and report metrics by model system before reporting all-model results.

Caveats

• This is a convenience table, not a single homogeneous model dataset.
• Random row-level splits can overstate performance.
• Report labels and non-event labels inherit all upstream report-matching assumptions.
• Use the separate HRRR, RAP, and GFS repos for cleaner single-system experiments.
• The union schema produces nulls where a model system does not provide a field. Nullness itself can become a model-system signal if splits are not designed carefully.

Attribution

This package is derived from public severe-weather reports and public numerical weather prediction data processed through a local rustwx severe-environment pipeline. Cite the original data providers and this dataset when using it in research or downstream model training.