Fine-tuned models repeat the context of questions.

[RuohuaTang]

Hi there,

After fine-tuning the model, I tried to make inferences. I strictly followed the data structure during fine-tuning. Specifically, my dataset contains "instruction", "input", and "output". However, when I make inferences using prompts like ["instruction: Correct the sentence. \ninput: Hellow World! \noutput: Hello World!"] or send the questions directly like ["Correct the sentence.\nHellow World!\n"], the model always repeats my questions and returns "instruction: Correct the sentence. \ninput: Hellow World! \noutput: Hello World!" or "Correct the sentence.\nHellow World!\nHello World!", respectively.

My questions are:

1. Will the generating process of repeating the context of questions use computing units (GPUs) and related sources? If it will, could you please let me know how to avoid this repeating?
2. Is there anything wrong of my understanding and setting of the data structure for fine-tuning? Because the raw model will not repeat the questions before being fine-tuned.

Many thanks in advance!