Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

finetune_version预测输出不可控的问题 #679

Open
elesun2018 opened this issue Dec 25, 2024 · 6 comments
Open

finetune_version预测输出不可控的问题 #679

elesun2018 opened this issue Dec 25, 2024 · 6 comments
Assignees

Comments

@elesun2018
Copy link

训练数据如下:
问题:图片中天气如何?
图片:图片路径
答案:不会下雨。

问题:图片中天气如何?
图片:图片路径
答案:会下雨。

答案只有会下雨和不会下雨。
训练样本约1000。
迭代训练5Kstep后,发现问:图片中天气如何?
答案:会下雨。但是大概20%的可能性会出现下述情况:
推理问题:图片中天气如何?
答案:图片中天气晴朗,没有乌云。。。。。描述了大量的通用文本,并没有输出我们想要的答案(不会下雨)。

请问问题原因可能是:训练不充分,数据量少,答案文本太短,需要多轮对话实现,微调参数设置lora?谢谢!

@zRzRzRzRzRzRzR zRzRzRzRzRzRzR self-assigned this Dec 30, 2024
@zRzRzRzRzRzRzR
Copy link
Member

100 step loss多少了

@elesun2018
Copy link
Author

image
batch4
100 step loss约为0.3
150 step后 loss趋于稳定0.2
请问这个step跟batch4 batch1有关系吧

@zRzRzRzRzRzRzR
Copy link
Member

没有太大关系,当然你batch越大鲁棒性繁华性越高。

@zRzRzRzRzRzRzR
Copy link
Member

现在这个loss应该能正常回答问题了,你是大概多少的相关数据呢,500-1000条吗

@elesun2018
Copy link
Author

elesun2018 commented Jan 2, 2025 via email

@zRzRzRzRzRzRzR
Copy link
Member

尝试进行更多轮训练,是否能实现相似效果,大概让loss降低到0.1,推理的时候保持贪婪采样

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

2 participants