weathernext2-mic

Slim model weights, executable usage and framework expectations in one repository.

• Contract manifest
• Transformers usage
• Transformers expectations
• Fixture provenance

Contract

WeatherNext2 forecast

A runnable contract for the companion test fixture. It exercises seeded synthetic physical inputs, real fixed geometry, batched forecasting, physical postprocessing, named gradients and an optimizer update using the official ensemble CRPS, checkpoint-specific loss weights and missing cyclone targets. Its synthetic fields are integration inputs, not a useful training dataset.

Variants cover default SDPA, eager attention and bfloat16. Numerical references are platform-specific; passing this contract does not establish forecast accuracy. See the fixture's build record for the manual construction and sensitivity-gate limitations.

Slim model

WeatherNext2 test fixture

A seeded random test artifact, not trained forecast weights. Derived from kashif/weathernext2-mini@36202f134c76a8e1e8259a82cdb437550935b390 using MIC's shrink recipe: width 64, intermediate width 128, four heads, two layers, and graph-edge width 64.

The fixture uses a 9-by-16 global grid, a 162-node spherical mesh and two-hop mesh attention. Its eight frozen geometry tensors and attention mask are regenerated together with the official checkpoint converter. Normalization, variable layouts and official mini loss weights are retained; the processor uses the new grid dimensions.

Run python rebuild.py OUTPUT with Transformers WeatherNext2 support and MIC at 97a594fdcd2610ecb2059853a2da9843f9d2149a. Geometry construction also requires SciPy, trimesh and rtree. This remains a manual MIC recipe: the stock builder does not regenerate WeatherNext2 geometry or preserve its attention-mask asset. Source pins, hashes and checks are in build.json.

The companion contract exercises batched forecasting, physical postprocessing, ensemble CRPS, named gradients and an optimizer update. Targets are synthetic test inputs.

The generic sensitivity gate has no perturbations for grid_features, global_features or noise and includes hidden states when comparing outputs. Passing it does not establish prediction sensitivity for every input or forecast accuracy.