diff --git a/model/model_training/custom_datasets/oasst_dataset.py b/model/model_training/custom_datasets/oasst_dataset.py
index 37bf2a71ca..eec0ce8849 100644
--- a/model/model_training/custom_datasets/oasst_dataset.py
+++ b/model/model_training/custom_datasets/oasst_dataset.py
@@ -98,6 +98,8 @@ def leaf_filter(thread: list[ExportMessageNode]) -> bool:
             elif mode == "rm":
                 # for reward models we use thread-fragments ending on prompter messages as prefix and
                 # their (ranked) replies as possible continuations.
+                if thread[-1].replies is None:
+                    return False
                 return (
                     thread[-1].role == "prompter"
                     and len([r for r in thread[-1].replies if r.rank is not None]) > 1