Update README with model loading code

README.md

@@ -15,13 +15,18 @@ Part of the [LAPVQA collection](https://huggingface.co/collections/dmusingu/lapv
 
 VQA task head trained on top of the **LAPVQA captioning-pretrained encoder**
 ([`lapvqa-pretrain-captioning`](https://huggingface.co/dmusingu/lapvqa-pretrain-captioning)).
-The encoder is kept frozen; this checkpoint contains the task head weights only.
+The encoder is kept frozen; this checkpoint contains the `VQAHead` state dict only.
+The encoder outputs 1024-dim patch tokens (ViT-L/14).
 
-This variant tests whether a generative pretraining objective on MIMIC-CXR
-produces better representations for VQA than off-the-shelf encoders.
+## Loading
 
-## Files
+```python
+import torch
+from lapvqa.vqa.model import VQAHead
 
-| File | Description |
-|---|---|
-| `pretrain-captioning_best.pt` | VQA head weights (encoder not included) |
+ckpt = torch.load("pretrain-captioning_best.pt", map_location="cpu")
+head = VQAHead(vis_dim=1024)
+head.load_state_dict(ckpt)
+head.eval()
+# pair with encoder_final.pt from lapvqa-pretrain-captioning
+```