diff --git a/src/datatrove/pipeline/inference/checkpointing.py b/src/datatrove/pipeline/inference/checkpointing.py
index 2160bdee..60ce9286 100644
--- a/src/datatrove/pipeline/inference/checkpointing.py
+++ b/src/datatrove/pipeline/inference/checkpointing.py
@@ -292,6 +292,7 @@ async def parse_existing_checkpoints(self, rank: int, output_writer_context: Dis
                 # not strictly needed but just to be safe for the future
                 async with self.file_locks[chunk_index]:
                     for document in reader.read_file(filename):
+                        document.metadata.pop("file_path", None)  # Remove any injected file_path
                         if "__no_rollouts_remove" not in document.metadata:
                             output_writer_context.write(document, rank=rank, chunk_index=chunk_index)
                         all_ids.add(document.id)