add retry loop around load_dataset for transient HF Hub 5xx

distill.py

@@ -293,7 +293,22 @@ class StreamingTextLoader:
         from datasets import load_dataset
         from datasets.distributed import split_dataset_by_node
 
-        ds = load_dataset(name, split="train", streaming=True)
+        # HF Hub occasionally returns 5xx during dataset metadata crawl. Retry.
+        last_err = None
+        for attempt in range(8):
+            try:
+                ds = load_dataset(name, split="train", streaming=True)
+                break
+            except Exception as e:
+                last_err = e
+                wait = min(2 ** attempt, 30)
+                log.warning(
+                    f"load_dataset({name!r}) failed (attempt {attempt + 1}/8): "
+                    f"{type(e).__name__}: {e}; sleeping {wait}s"
+                )
+                time.sleep(wait)
+        else:
+            raise RuntimeError(f"load_dataset failed after 8 retries") from last_err
         ds = ds.shuffle(seed=seed, buffer_size=shuffle_buffer)
         ds = split_dataset_by_node(ds, rank=rank, world_size=world_size)
         self._ds = iter(ds)