README.md

---
license: apache-2.0
library_name: pytorch
tags:
- weather
- multimodal
- polaris
- component-extraction
base_model: sais-org/MKB
---
# MKB weather encoder, identity projector and decoder

Native Polaris/Swin weather modules from the official MKB unified scientific checkpoint. All selected parameters retain their original names, dtypes and exact values. The weather encoder maps meteorological fields to 4096-dimensional language-backbone features; the decoder maps language hidden states and the encoder's shared step cache back to dense fields.

```python
from huggingface_hub import snapshot_download
import sys
folder=snapshot_download("gatilin/mkb_weather_encoder_decoder")
sys.path.insert(0,folder)
from component import load_model
weather=load_model(folder) # default native geometry 721x1440
router=weather.model.modality_router
encoder=router.encoders["weather"]
decoder=router.decoders["weather"]
```

Use the original MKB task protocol, ERA5 variable ordering, input normalization, static fields and MKB language backbone for meaningful physical forecasts. Encoder and decoder must remain wired together: the decoder reuses condition embeddings, post-Swin skip features, meteorological inputs and grid metadata from the encoder cache. The projector is intentionally a no-op with its original `_noop` checkpoint entry.

## Validation scope

`python validate.py` strictly loads all 299 selected tensor entries and executes the native encoder and decoder at a **120x120 spatial geometry override** using all original learned dimensions and parameters. It tests the encoder's 400x4096 features, identity projector, cached encoder/decoder interface and float32 residual prediction. Inputs are synthetic standardized 70-channel fields, zero static constants, and explicitly synthetic language hidden states. The test does **not** run the 721x1440 global model, evaluate ERA5 forecast skill, perform physical de-normalization or a multi-day rollout. Complete settings are in validation.json.

The full MKB language backbone, biological branches and medical SAM3 branch are excluded. This package therefore does not claim complete scientific-model capabilities or inherit a fake blanket license for SAM3; only the Apache-2.0 weather component is distributed here. Source revision and exact tensor mappings are recorded in the extraction and provenance files.

Sources: [official MKB checkpoint](https://huggingface.co/sais-org/MKB), [code](https://github.com/Shanghai-Academy-of-AI-For-Science/MKB), [paper](https://arxiv.org/abs/2607.20557).