diff --git a/include/onnxruntime/core/optimizer/graph_transformer_config.h b/include/onnxruntime/core/optimizer/graph_transformer_config.h
index 42395ff303..6866433d51 100644
--- a/include/onnxruntime/core/optimizer/graph_transformer_config.h
+++ b/include/onnxruntime/core/optimizer/graph_transformer_config.h
@@ -51,7 +51,7 @@ struct GraphTransformerConfiguration {
       return (s1 == s2) == false;
     }
 
-    int level{-1}; /* -1 => no cast propagation,
+    int level{1}; /* -1 => no cast propagation,
                        0 => use user specified list of opcodes to allow moving cast operations,
                        1 => use ORT predefined list of level 1 opcodes in addition to the user specified allow opcodes
                        2 => use ORT predefined list of level 2 opcodes in addition to the user specified allow opcodes
diff --git a/orttraining/orttraining/python/orttraining_pybind_state.cc b/orttraining/orttraining/python/orttraining_pybind_state.cc
index 31c61727a8..a7eabe9b4b 100644
--- a/orttraining/orttraining/python/orttraining_pybind_state.cc
+++ b/orttraining/orttraining/python/orttraining_pybind_state.cc
@@ -71,7 +71,7 @@ struct TrainingParameters {
   bool enable_adasum = false;
 
   // transformation
-  int propagate_cast_ops_level = -1;
+  int propagate_cast_ops_level = 1;
   std::vector<std::string> propagate_cast_ops_allow;
   GraphTransformerConfiguration::PropagateCastOpsConfiguration::Strategy propagate_cast_ops_strategy =
       GraphTransformerConfiguration::PropagateCastOpsConfiguration::Strategy::None;