bert performance

[flyoliver]

hi, i finetune a classify model based on the original bert model and this repo. accuracy dropped 1%. parameters:
nohup python run_classifier.py --task_name=sex --do_train=true --do_eval=true --data_dir=** --vocab_file=** --bert_config_file=** --init_checkpoint=** --max_seq_length=128 --train_batch_size=16 --learning_rate=4e-5 --num_train_epochs=5 --output_dir=** --num_gpus=8 &
is my parameters wrong?