diff --git a/olmo/data/memmap_dataset.py b/olmo/data/memmap_dataset.py
index 879cfa761..b4e77c3bb 100644
--- a/olmo/data/memmap_dataset.py
+++ b/olmo/data/memmap_dataset.py
@@ -8,7 +8,7 @@
 from torch.utils.data import Dataset
 
 from ..aliases import PathOrStr
-from ..util import file_size, get_bytes_range
+from ..util import _get_s3_client, file_size, get_bytes_range
 
 __all__ = ["MemMapDataset"]
 
@@ -70,6 +70,9 @@ def max_seq_len(self) -> int:
 
     @property
     def offsets(self) -> List[Tuple[int, int]]:
+        # Create the global S3 client up front to work around a threading issue in boto.
+        _get_s3_client()
+
         if self._mmap_offsets is None:
             import concurrent.futures