推荐一篇讲decoding搜索策略的文章,应该可以通过设置decoding参数解决issue里很多输出不停的问题 #525
sunyuhan19981208
started this conversation in
General
Replies: 1 comment
-
但temperature过高也有坏处,可能会导致对事实性问题回答的准确率降低。 |
Beta Was this translation helpful? Give feedback.
0 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
https://towardsdatascience.com/the-three-decoding-methods-for-nlp-23ca59cb1e9d
就是这篇文章,主要就是将了大模型也会用到的三种decoding搜索策略,issue里面很多人都提到了输出不停,循环输出的问题,那么通过设置temperature提高random sampling的使用概率我觉得就有可能能解决这一问题,无论是greedy decoding还是beam search都会导向固定的结果,就有可能出现循环,random sampling是在多种可能的token里按置信度概率去随机选下一个token,这样就可能跳出循环
Beta Was this translation helpful? Give feedback.
All reactions