diff --git a/README.md b/README.md
index dbaefdd7..74182bc0 100644
--- a/README.md
+++ b/README.md
@@ -653,7 +653,7 @@ BQ Corpus (Bank Question corpus) is a Chinese corpus for sentence semantic equiv
 
 ### Install PaddlePaddle
 
-This code base has been tested with Paddle Fluid 1.6.3 with Python 2/3.5+, since Paddle 1.6.3 has changed some of APIs, using version before 1.6.3 might have bug on NER tasks.
+This code base has been tested with Paddle Fluid 1.6 with Python 2/3.5+, since Paddle 1.6 has changed some of APIs, using version before 1.6 might have bug on NER tasks.
 
 **\*Important\*** When finished installing Paddle Fluid, remember to update LD_LIBRARY_PATH about CUDA, cuDNN, NCCL2, for more information on paddlepaddle setup, you can click [here](http://en.paddlepaddle.org/documentation/docs/en/1.5/beginners_guide/index_en.html) and [here](http://en.paddlepaddle.org/documentation/docs/en/1.5/beginners_guide/install/install_Ubuntu_en.html). Also, you can read FAQ at the end of this document when you encounter errors.
 
diff --git a/README.zh.md b/README.zh.md
index 6008d1a3..9fce0f3d 100644
--- a/README.zh.md
+++ b/README.zh.md
@@ -689,7 +689,7 @@ ERNIE Tiny轻量级模型在公开数据集的效果如下所示，任务均值
 
 ## PaddlePaddle安装
 
-本项目依赖于 Paddle 1.6.3，* 由于Paddle 1.6.3版本相比之前版本有较大API改动，使用Paddle 1.6.3以前版本运行本代码库会导致序列标注等任务报错 *，请参考[安装指南](http://www.paddlepaddle.org/#quick-start)进行安装。
+本项目依赖于 Paddle 1.6，* 由于Paddle 1.6版本相比之前版本有较大API改动，使用Paddle 1.6以前版本运行本代码库会导致序列标注等任务报错 *，请参考[安装指南](http://www.paddlepaddle.org/#quick-start)进行安装。
 
 **【重要】安装后，需要及时的将 CUDA、cuDNN、NCCL2 等动态库路径加入到环境变量 LD_LIBRARY_PATH 之中，否则训练过程中会报相关的库错误。具体的paddlepaddle配置细节请查阅[这里](http://en.paddlepaddle.org/documentation/docs/zh/1.5/beginners_guide/quick_start_cn.html)**
 
diff --git a/ernie/finetune/mrc.py b/ernie/finetune/mrc.py
index 3cc13a86..cfe3bbdb 100644
--- a/ernie/finetune/mrc.py
+++ b/ernie/finetune/mrc.py
@@ -44,7 +44,7 @@ def create_model(args, pyreader_name, ernie_config, is_training):
     pos_ids = fluid.layers.data(name='2', shape=[-1, args.max_seq_len, 1], dtype='int64')
     sent_ids= fluid.layers.data(name='3', shape=[-1, args.max_seq_len, 1], dtype='int64')
     task_ids= fluid.layers.data(name='4', shape=[-1, args.max_seq_len, 1], dtype='int64')
-    input_mask = fluid.layers.data(name='5', shape=[-1, 1], dtype='float32')
+    input_mask = fluid.layers.data(name='5', shape=[-1, args.max_seq_len, 1], dtype='float32')
     start_positions = fluid.layers.data(name='6', shape=[-1, 1], dtype='int64')
     end_positions = fluid.layers.data(name='7', shape=[-1, 1], dtype='int64')
     unique_id = fluid.layers.data(name='8', shape=[-1, 1], dtype='int64')
diff --git a/ernie/run_mrc.py b/ernie/run_mrc.py
index b8b4e42d..4181e45f 100644
--- a/ernie/run_mrc.py
+++ b/ernie/run_mrc.py
@@ -361,7 +361,6 @@ def main(args):
 if __name__ == '__main__':
     prepare_logger(log)
     print_arguments(args)
-    while True:
-        scope = fluid.core.Scope()
-        with fluid.scope_guard(scope):
-            main(args)
+    scope = fluid.core.Scope()
+    with fluid.scope_guard(scope):
+        main(args)