diff --git a/returnn/datasets/postprocessing.py b/returnn/datasets/postprocessing.py
index 738c87412..08cac3c55 100644
--- a/returnn/datasets/postprocessing.py
+++ b/returnn/datasets/postprocessing.py
@@ -364,11 +364,11 @@ def __call__(self, iterator: Iterator[TensorDict], **kwargs) -> Iterator[TensorD
         is_down_phase = False
 
         seq_buffer = list(islice(iterator, self.num_seqs_per_bin))
+        has_ended = False
         while True:
             seq_buffer.sort(key=self._get_seq_len, reverse=is_down_phase)
 
             next_seq_buffer = []
-            has_ended = False
 
             # Yield items to trainer while gradually pulling more data from PP function.
             # This optimizes CPU load when multiple workers are used.