Add 'Getting the weights' note; align with GitHub mirror

.gitignore

@@ -6,7 +6,13 @@ __pycache__/
 dist/
 build/
 .env
-# norm_stats and static_vars .npz are bundled release assets — track via LFS (see .gitattributes)
-# generated forecast outputs only:
+
+# Model checkpoints are not stored in git — fetch from Hugging Face
+# (huggingface.co/maxxxzdn/mosaic) instead. On HF they live as LFS objects.
+*.pt
+*.pth
+*.ckpt
+
+# Generated forecast outputs (kept out of git but bundled .npz assets are tracked)
 verify_*.npz
 forecast_*.npz

README.md

@@ -104,6 +104,24 @@ For reading data from Google Cloud Storage (WeatherBench2 zarr stores):
 pip install gcsfs
 ```
 
+## Getting the weights
+
+If you installed via Hugging Face (`huggingface-cli download maxxxzdn/mosaic --local-dir .`), the checkpoints (`era5_best.pt`, `hres_best.pt`) and normalization stats are already bundled and `inference.py` finds them in the working directory.
+
+If you cloned this repo from GitHub instead, the weights are not in the git tree (they live on the [Hugging Face mirror](https://huggingface.co/maxxxzdn/mosaic) as LFS objects). Fetch them with:
+
+```bash
+pip install huggingface_hub
+huggingface-cli download maxxxzdn/mosaic --local-dir .   # pulls .pt + .npz assets
+```
+
+or programmatically:
+
+```python
+from huggingface_hub import snapshot_download
+snapshot_download(repo_id="maxxxzdn/mosaic", local_dir=".")
+```
+
 ## Quick Start
 
 ```bash