---
license: mit
library_name: pytorch
tags:
  - weather
  - data-assimilation
  - diffusion-models
  - aurora
---

# SBDA Resources

This repository contains the model checkpoints, normalization statistics, and
sparse-observation mask required to reproduce the SBDA experiments.

## Files

| Path | Purpose |
| --- | --- |
| `model_ckpt/sbda_full.pt` | Full SBDA checkpoint |
| `model_ckpt/sbda_wo_oi_train.pt` | SBDA checkpoint trained without observation refinement |
| `model_ckpt/diffda.pt` | Baseline DiffDA checkpoint |
| `aifs_ckpt/aurora-1.5-finetuned.pt` | Fine-tuned Aurora checkpoint for background generation and forecast evaluation |
| `stats/mean_std_121x240.npz` | ERA5 normalization statistics used by all experiments |
| `stats/mean_std_diff_120x240.npz` | ERA5–Aurora difference statistics used by baseline DiffDA |
| `mask/mask_120x240_1pct.pt` | Fixed 1% sparse-observation mask |

Download all resources into the `resource/` directory of the code release:

```bash
hf download strayee/SBDA --local-dir resource
```

File integrity can be checked with:

```bash
sha256sum -c SHA256SUMS
```

## License

The distributed checkpoints are released under the MIT terms in `LICENSE.md`.
The fine-tuned Aurora checkpoint retains the Microsoft Aurora MIT notice. The
license of these files does not relicense the software required to run them or
the ERA5-derived source data; see `LICENSE.md` for the precise scope.

PyTorch `.pt` files may use Python pickle internally. Load checkpoint files only
from a source you trust.