readme: reference .ptw weights

README.md

@@ -69,8 +69,8 @@ Trained PIVOT checkpoints live under `models/`:
 - `models/norman/` - trained on Norman 2019 (CRISPRa K562)
 - `models/replogle_k562/` - trained on Replogle 2022 (CRISPRi K562)
 
-each folder has `model.pt` (state dict), `config.json` (the training config), and
-`train_info.json` (history + run info). loading needs the matching preprocessed dataset,
+each folder has `model.ptw` (a plain torch state dict), `config.json` (the training config),
+and `train_info.json` (history + run info). loading needs the matching preprocessed dataset,
 since the perturbation encoder vocabulary comes from the data:
 
 ```python
@@ -81,7 +81,7 @@ from src.training.train import TrainConfig, make_model
 cfg = TrainConfig(**json.load(open("models/norman/config.json")))
 data = load_dataset(cfg.dataset)
 model = make_model(data, cfg, device="cpu")
-model.load_state_dict(torch.load("models/norman/model.pt", map_location="cpu"))
+model.load_state_dict(torch.load("models/norman/model.ptw", map_location="cpu"))
 model.eval()
 ```