diff --git a/examples/bert/bert_classifier.py b/examples/bert/bert_classifier.py
index 472e39b3436bfe9c32a6941bcfeeac6badb82e5f..e3c9726c9a1eb257748e7a8d2fb35e316b7bd80b 100644
--- a/examples/bert/bert_classifier.py
+++ b/examples/bert/bert_classifier.py
@@ -157,7 +157,7 @@ def main():
         labels,
         device=device)
 
-    cls_model.bert_layer.load("./bert_small", reset_optimizer=True)
+    cls_model.bert_layer.load("./bert_uncased_L-12_H-768_A-12/bert", reset_optimizer=True)
 
     # do train
     cls_model.fit(train_data=train_dataloader.dataloader,
diff --git a/examples/bert/readme.md b/examples/bert/readme.md
new file mode 100644
index 0000000000000000000000000000000000000000..f8df579ec8302b890f7fb523facc417fc784536a
--- /dev/null
+++ b/examples/bert/readme.md
@@ -0,0 +1,9 @@
+1. download data: wget https://paddle-hapi.bj.bcebos.com/data/bert_data.tar.gz
+
+2. unzip data: tar -zvxf bert_data.tar.gz
+
+3. download pretrained parameters: wget https://paddle-hapi.bj.bcebos.com/models/bert_uncased_L-12_H-768_A-12.tar.gz
+
+4. unzip pretrained parameters: tar -zvxf bert_uncased_L-12_H-768_A-12.tar.gz
+
+4. bash run_classifier_single_gpu.sh 
diff --git a/examples/bert/run_classifier_multi_gpu.sh b/examples/bert/run_classifier_multi_gpu.sh
new file mode 100755
index 0000000000000000000000000000000000000000..c3800568da786d1dec170d74022443d65f175f66
--- /dev/null
+++ b/examples/bert/run_classifier_multi_gpu.sh
@@ -0,0 +1,25 @@
+#!/bin/bash
+BERT_BASE_PATH="./bert_uncased_L-12_H-768_A-12/"
+DATA_PATH="./data/glue_data/MNLI/"
+CKPT_PATH="./data/saved_model/mnli_models"
+
+# start fine-tuning
+python3.7 -m paddle.distributed.launch --started_port 8899 --selected_gpus=1,2,3 bert_classifier.py\
+    --use_cuda true \
+    --do_train true \
+    --do_test true \
+    --batch_size 64 \
+    --data_dir ${DATA_PATH} \
+    --vocab_path ${BERT_BASE_PATH}/vocab.txt \
+    --checkpoints ${CKPT_PATH} \
+    --save_steps 1000 \
+    --weight_decay  0.01 \
+    --warmup_proportion 0.1 \
+    --validation_steps 100 \
+    --epoch 3 \
+    --max_seq_len 128 \
+    --bert_config_path ${BERT_BASE_PATH}/bert_config.json \
+    --learning_rate 5e-5 \
+    --skip_steps 10 \
+    --shuffle true
+