STF后发现模型的基础能力丢失
#5733
Replies: 2 comments 2 replies
-
+1 SFT qwen2.5后发现 system prompt的灵活感知能力没有了。qwen2没有这种情况 |
Beta Was this translation helpful? Give feedback.
2 replies
-
确认是模型问题,不是 llama-factory 的坑 |
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
-
Reminder
System Info
用一波客服的对话数据做SFT(Qwen2.5-0.5B-Instruct),LoRA微调后(默认参数,最终loss=2.0),发现模型的一部分基础能力没了(比如system prompt: 用英文回答所有问题,实际回复还是中文,base模型没问题的),而换一个对话数据集就没事,确认可能是数据集有问题,但目测都是很正常的对话,完全看不出区别。有人遇到过类似问题么?
Reproduction
Expected behavior
No response
Others
No response
Beta Was this translation helpful? Give feedback.
All reactions