Add format section explaining safetensors compatibility

README.md

@@ -66,14 +66,23 @@ Pre-trained [TabDPT](https://github.com/JesseCresswell/tfm-mia) model weights tr
 | CTR Correlation | 0.830 | 0.830 | 0.827 |
 | CTR R² | 0.726 | 0.730 | 0.725 |
 
+## Format
+
+These checkpoints were converted from PyTorch Lightning `.ckpt` files (which include optimizer state, ~295MB each) to SafeTensors format (model weights only, ~103MB each). This is the same format used by the official `Layer6/TabDPT` release. The `tabdpt` package natively loads SafeTensors via the `model_weight_path` argument — no extra conversion needed.
+
 ## Usage
 
 ```python
 from tabdpt import TabDPTClassifier
 from huggingface_hub import hf_hub_download
 
+# Download once (cached afterwards)
 path = hf_hub_download("dwahdany/TabDPT", "production_seed42.safetensors")
+
+# Use exactly like the default model
 clf = TabDPTClassifier(model_weight_path=path)
 clf.fit(X_train, y_train)
 preds = clf.predict(X_test)
 ```
+
+Works identically with `TabDPTRegressor`.