Firefly-LLaMA2-Chinese icon indicating copy to clipboard operation
Firefly-LLaMA2-Chinese copied to clipboard

Published 20 hours ago • verified publisher icon yangjianxin1

baichaun2-13b增量预训练loss为0

Open LiuChen19960902 opened this issue 1 year ago • 1 comments

作者你好，我使用baichuan2-13b做增量cpt时候loss一直是0. 我使用自己的数据集或是CNEWsum.jsonl都是0.

image

LiuChen19960902 avatar Sep 26 '23 06:09 LiuChen19960902

真没搞懂作者训练baichuan的作用是啥

xxllp avatar Sep 26 '23 15:09 xxllp