diff --git a/convlab2/base_models/t5/nlu/run_nlu.sh b/convlab2/base_models/t5/nlu/run_nlu.sh
index 71b5f5f8e488b2c44e76637cc0152b8ff9db8671..05671139cfd691422c924c9da880af317c916a19 100644
--- a/convlab2/base_models/t5/nlu/run_nlu.sh
+++ b/convlab2/base_models/t5/nlu/run_nlu.sh
@@ -40,6 +40,7 @@ python ../run_seq2seq.py \
     --do_eval \
     --save_strategy epoch \
     --evaluation_strategy epoch \
+    --save_total_limit 3 \
     --prediction_loss_only \
     --cache_dir ${cache_dir} \
     --output_dir ${output_dir} \
@@ -72,7 +73,14 @@ python ../run_seq2seq.py \
     --logging_dir ${logging_dir} \
     --overwrite_output_dir \
     --preprocessing_num_workers 4 \
-    --per_device_eval_batch_size ${per_device_eval_batch_size}
+    --per_device_train_batch_size ${per_device_train_batch_size} \
+    --per_device_eval_batch_size ${per_device_eval_batch_size} \
+    --gradient_accumulation_steps ${gradient_accumulation_steps} \
+    --learning_rate ${lr} \
+    --num_train_epochs ${num_train_epochs} \
+    --debug underflow_overflow \
+    --adafactor \
+    --gradient_checkpointing
 
 python merge_predict_res.py -d ${dataset_name} -s ${speaker} -c ${context_window_size} -p ${output_dir}/generated_predictions.json
 
diff --git a/convlab2/base_models/t5/nlu/run_nlu_fewshot.sh b/convlab2/base_models/t5/nlu/run_nlu_fewshot.sh
index 026e50aa68a7e92ed8294e73f6a5d0c531d59838..2c783912d66281c8e44f01aafc232b4051e73f86 100644
--- a/convlab2/base_models/t5/nlu/run_nlu_fewshot.sh
+++ b/convlab2/base_models/t5/nlu/run_nlu_fewshot.sh
@@ -76,7 +76,14 @@ python ../run_seq2seq.py \
     --logging_dir ${logging_dir} \
     --overwrite_output_dir \
     --preprocessing_num_workers 4 \
-    --per_device_eval_batch_size ${per_device_eval_batch_size}
+    --per_device_train_batch_size ${per_device_train_batch_size} \
+    --per_device_eval_batch_size ${per_device_eval_batch_size} \
+    --gradient_accumulation_steps ${gradient_accumulation_steps} \
+    --learning_rate ${lr} \
+    --num_train_epochs ${num_train_epochs} \
+    --debug underflow_overflow \
+    --adafactor \
+    --gradient_checkpointing
 
 python merge_predict_res.py -d ${dataset_name} -s ${speaker} -c ${context_window_size} -p ${output_dir}/generated_predictions.json -o ${dial_ids_order}
 
diff --git a/convlab2/base_models/t5/nlu/run_nlu_pretrain.sh b/convlab2/base_models/t5/nlu/run_nlu_pretrain.sh
new file mode 100644
index 0000000000000000000000000000000000000000..3a059c2d27c809a3a1cf26b3a49160cab96dd1d4
--- /dev/null
+++ b/convlab2/base_models/t5/nlu/run_nlu_pretrain.sh
@@ -0,0 +1,67 @@
+n_gpus=1
+task_name="nlu"
+dataset_name="sgd+tm1+tm2+tm3"
+speaker="user"
+context_window_size=0
+data_dir="data/${task_name}/${dataset_name}/${speaker}/context_${context_window_size}"
+output_dir="output/${task_name}/${dataset_name}/${speaker}/context_${context_window_size}"
+cache_dir="../cache"
+logging_dir="${output_dir}/runs"
+train_file="${data_dir}/train.json"
+validation_file="${data_dir}/validation.json"
+test_file="${data_dir}/test.json"
+metric_name_or_path="nlu_metric.py"
+metric_for_best_model="overall_f1"
+source_column="context"
+target_column="dialogue_acts_seq"
+truncation_side="left"
+max_source_length=512
+max_target_length=512
+model_name_or_path="t5-small"
+per_device_train_batch_size=128
+per_device_eval_batch_size=64
+gradient_accumulation_steps=2
+lr=1e-3
+num_train_epochs=1
+
+names=$(echo ${dataset_name} | tr "+" "\n")
+mkdir -p ${data_dir}
+for name in ${names};
+do
+    echo "preprocessing ${name}"
+    python ../create_data.py -t ${task_name} -d ${name} -s ${speaker} -c ${context_window_size}
+    if [ "${name}" != "${dataset_name}" ]; then
+        cat "data/${task_name}/${name}/${speaker}/context_${context_window_size}/train.json" >> ${train_file}
+        cat "data/${task_name}/${name}/${speaker}/context_${context_window_size}/validation.json" >> ${validation_file}
+        cat "data/${task_name}/${name}/${speaker}/context_${context_window_size}/test.json" >> ${test_file}
+    fi
+done
+
+python ../run_seq2seq.py \
+    --task_name ${task_name} \
+    --train_file ${train_file} \
+    --validation_file ${validation_file} \
+    --source_column ${source_column} \
+    --target_column ${target_column} \
+    --max_source_length ${max_source_length} \
+    --max_target_length ${max_target_length} \
+    --truncation_side ${truncation_side} \
+    --model_name_or_path ${model_name_or_path} \
+    --do_train \
+    --do_eval \
+    --save_strategy epoch \
+    --evaluation_strategy epoch \
+    --prediction_loss_only \
+    --cache_dir ${cache_dir} \
+    --output_dir ${output_dir} \
+    --logging_dir ${logging_dir} \
+    --overwrite_output_dir \
+    --preprocessing_num_workers 4 \
+    --per_device_train_batch_size ${per_device_train_batch_size} \
+    --per_device_eval_batch_size ${per_device_eval_batch_size} \
+    --gradient_accumulation_steps ${gradient_accumulation_steps} \
+    --learning_rate ${lr} \
+    --num_train_epochs ${num_train_epochs} \
+    --debug underflow_overflow \
+    --adafactor \
+    --gradient_checkpointing