diff --git a/notebooks/laya_finetune_typed_decisions_mps.py b/notebooks/laya_finetune_typed_decisions_mps.py
index 4403334..ff465f4 100644
--- a/notebooks/laya_finetune_typed_decisions_mps.py
+++ b/notebooks/laya_finetune_typed_decisions_mps.py
@@ -230,6 +230,9 @@ def train(args, model_dir, items_path, device):
     n_calib = min(args.calib_max, len(all_items) // 10)
     calib_items = [all_items[i] for i in sorted(order[:n_calib])]
     train_items = [all_items[i] for i in sorted(order[n_calib:])]
+    if args.calib_items:
+        calib_items = torch.load(args.calib_items, map_location="cpu", weights_only=False)
+        train_items = all_items
 
     encoder_params = [p for n, p in model.named_parameters() if "encoder." in n]
     head_params = [p for n, p in model.named_parameters() if "encoder." not in n]
@@ -348,6 +351,7 @@ def main():
     parser.add_argument("--micro-batch", type=int, default=2)
     parser.add_argument("--grad-accum", type=int, default=16)
     parser.add_argument("--calib-max", type=int, default=400)
+    parser.add_argument("--calib-items", help="held-out items .pt used only for temperature; trains on all --items")
     parser.add_argument("--device", choices=["auto", "mps", "cpu"], default="auto")
     parser.add_argument("--force-preprocess", action="store_true")
     parser.add_argument("--no-checkpointing", action="store_true")