diff --git a/open_diloco/train_diloco_torch.py b/open_diloco/train_diloco_torch.py
index 6fa66d0..cb68c0f 100644
--- a/open_diloco/train_diloco_torch.py
+++ b/open_diloco/train_diloco_torch.py
@@ -230,6 +230,10 @@ def tokenize_function(data):
 
     if eval_steps is not None:
         eval_dataset = tokenized_datasets["validation"]
+        # Limit streaming validation to 1000 samples to avoid infinite loop
+        # (IterableDataset has no __len__, causing DataLoader to loop forever)
+        if hasattr(eval_dataset, "take"):
+            eval_dataset = eval_dataset.take(1000)
         eval_dataloader = DataLoader(
             eval_dataset,
             collate_fn=data_collator,