想问一下在指令微调时，是继续在原lora权重基础上训，还是重新训练新的权重

[SupritYoung]

作者有对比过二者的区别吗？

[lywang76]

作者有对比过二者的区别吗？

i just found that in finetune.sh, set the following --resume_from_checkpoint "./outputs/train_clm/"

After that, you can continuously train from trained lora weights.