diff --git a/egs/librispeech/ASR/pruned_transducer_stateless_gtrans/.train.py.swp b/egs/librispeech/ASR/pruned_transducer_stateless_gtrans/.train.py.swp
index 560db54c3..9f30cb5c4 100644
Binary files a/egs/librispeech/ASR/pruned_transducer_stateless_gtrans/.train.py.swp and b/egs/librispeech/ASR/pruned_transducer_stateless_gtrans/.train.py.swp differ
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless_gtrans/train.py b/egs/librispeech/ASR/pruned_transducer_stateless_gtrans/train.py
index 2531e14aa..336855bad 100755
--- a/egs/librispeech/ASR/pruned_transducer_stateless_gtrans/train.py
+++ b/egs/librispeech/ASR/pruned_transducer_stateless_gtrans/train.py
@@ -174,8 +174,7 @@ def add_model_arguments(parser: argparse.ArgumentParser):
     parser.add_argument(
         "--group-size",
         type=int,
-        default=4,
-        help="How many left context can be seen in chunks when calculating attention.",
+        default=12,
     )