diff --git a/model_zoo/official/nlp/bert/README.md b/model_zoo/official/nlp/bert/README.md
index 48423e32f7..f8f7568fb6 100644
--- a/model_zoo/official/nlp/bert/README.md
+++ b/model_zoo/official/nlp/bert/README.md
@@ -381,6 +381,11 @@ epoch: 0.0, current epoch percent: 0.000, step: 2, outpus are (Tensor(shape=[1],
 > ```
 > This will extend the timeout limits of hccl from the default 120 seconds to 600 seconds.
 
+> **Attention** If you are running with a big bert model, some error of protobuf may occurs while saving checkpoints, try with the following environ set.
+> ```
+> export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
+> ```
+
 ### Distributed Training
 #### Running on Ascend
 ```
diff --git a/model_zoo/official/nlp/bert/src/utils.py b/model_zoo/official/nlp/bert/src/utils.py
index 77c71d2b88..6d6a6196b1 100644
--- a/model_zoo/official/nlp/bert/src/utils.py
+++ b/model_zoo/official/nlp/bert/src/utils.py
@@ -145,11 +145,17 @@ class LossCallBack(Callback):
         super(LossCallBack, self).__init__()
         self._dataset_size = dataset_size
     def step_end(self, run_context):
+        """
+        Print loss after each step
+        """
         cb_params = run_context.original_args()
         if self._dataset_size > 0:
             percent, epoch_num = math.modf(cb_params.cur_step_num / self._dataset_size)
+            if percent == 0:
+                percent = 1
+                epoch_num -= 1
             print("epoch: {}, current epoch percent: {}, step: {}, outputs are {}"
-                  .format(epoch_num, "%.3f" % percent, cb_params.cur_step_num, str(cb_params.net_outputs)))
+                  .format(int(epoch_num), "%.3f" % percent, cb_params.cur_step_num, str(cb_params.net_outputs)))
         else:
             print("epoch: {}, step: {}, outputs are {}".format(cb_params.cur_epoch_num, cb_params.cur_step_num,
                                                                str(cb_params.net_outputs)))