diff --git a/src/llmtuner/tuner/sft/workflow.py b/src/llmtuner/tuner/sft/workflow.py
index 05942780..d45571d2 100644
--- a/src/llmtuner/tuner/sft/workflow.py
+++ b/src/llmtuner/tuner/sft/workflow.py
@@ -39,7 +39,7 @@ def run_sft(
     # Override the decoding parameters of Seq2SeqTrainer
     training_args_dict = training_args.to_dict()
     training_args_dict.update(dict(
-        generation_max_length=training_args.generation_max_length or data_args.max_target_length,
+        generation_max_length=training_args.generation_max_length or data_args.cutoff_len,
         generation_num_beams=data_args.eval_num_beams or training_args.generation_num_beams
     ))
     training_args = Seq2SeqTrainingArguments(**training_args_dict)