Replies: 19 comments
-
你用的p-tuning吗?还是lora? |
Beta Was this translation helpful? Give feedback.
-
|
Beta Was this translation helpful? Give feedback.
-
我也是这个问题 减少文本长度有用么? |
Beta Was this translation helpful? Give feedback.
-
我在alpaca-lora遇到了同样的问题,猜测生成长文本的难度确实较大,而且lora在微调方法中也算是属于效果不太好的那一类了 |
Beta Was this translation helpful? Give feedback.
-
我用p-tuning也是出现这种 |
Beta Was this translation helpful? Give feedback.
-
max_target_length别指定太长,缩短到64,训练次数适当增加。部署时,tockenizer就用原本的Thudm/ChatGLM-6B,model用ptuning的model,你这情况我在训练llama时遇到过。chatglm没遇见。
发自我的iPhone
… 在 2023年4月20日,10:22,Leawnn ***@***.***> 写道:
我用p-tuning也是出现这种
求大神解答
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you commented.
|
Beta Was this translation helpful? Give feedback.
-
max_target_length为128时,生成中文文本字数大概为200字左右,你这个max_target_length=300,我猜测如果基底模型生成的内容不够长的话,就会不停重复 |
Beta Was this translation helpful? Give feedback.
-
同样遇到了,有时候会疯狂排比句 |
Beta Was this translation helpful? Give feedback.
-
用全量微调 感觉没有这种问题了 |
Beta Was this translation helpful? Give feedback.
-
@liuanping 大佬解决了吗,我也出现了同样的问题 |
Beta Was this translation helpful? Give feedback.
-
感觉全量微调就好了 |
Beta Was this translation helpful? Give feedback.
-
@Lufffya 感觉全量微调就好了 还有一些办法说是加重复惩罚项 |
Beta Was this translation helpful? Give feedback.
-
哦哦好吧,那估计不太行,我这里只有一张4090,跑不起来,谢谢 |
Beta Was this translation helpful? Give feedback.
-
@Lufffya 可能chatuan更友好 因为他是10亿参数模型 效果也还行。 |
Beta Was this translation helpful? Give feedback.
-
请问你有全量微调嘛 |
Beta Was this translation helpful? Give feedback.
-
@liuanping 大佬解决了吗,我也出现了同样的问题 |
Beta Was this translation helpful? Give feedback.
-
@shuanglong520 我全量微调的时候发现没了 lora不行可能 |
Beta Was this translation helpful? Give feedback.
-
Is there an existing issue for this?
Current Behavior
我发现长文本生成效果不容易调出来,特别容易出现重复 例如“我吃饭了吗吗吗吗吗吗吗吗吗”,网上说是退化问题,即随着生成文本长度的增加其质量会逐渐降低,容易出现多种层次(字、短语、句子级)的重复生成。有没有大神给一些有效的经验。
Expected Behavior
求大神指点
Steps To Reproduce
训练长文本生成。
Environment
Anything else?
none
Beta Was this translation helpful? Give feedback.
All reactions