Qwen3训练非思考数据,数据集是否label是否可以不带/think标签,如果数据集全部不带/think标签是否会影响模型原本的能力 ? #1510
Unanswered
shenshaowei
asked this question in
Q&A
Replies: 3 comments 4 replies
-
有解决方法了吗大佬? |
Beta Was this translation helpful? Give feedback.
2 replies
-
可带可不带,默认就是思考模式
…----------回复的邮件信息----------
DojiSum ***@***.***> 在 2025年7月28日 15:14写道:
我们也遇到这个问题了,思路和你的挺像
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you authored the thread.Message ID: ***@***.***>
|
Beta Was this translation helpful? Give feedback.
2 replies
-
混合少量思考数据即可
…----------回复的邮件信息----------
DojiSum ***@***.***> 在 2025年7月28日 15:46写道:
嗯嗯,我用/no_think的数据训,测评时发现会丢失思考。你有什么缓解的方法吗? :)
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you authored the thread.Message ID: ***@***.***>
|
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
Uh oh!
There was an error while loading. Please reload this page.
-
Qwen3训练非思考数据,数据集是否label是否可以不带/think标签,如果数据集全部不带/think标签是否会影响模型原本的能力 ,我训练发现思考能力丢失了,https://github.com/modelscope/ms-swift/issues/4030的格式:
方案一:在训练时额外指定--loss_scale ignore_empty_think,忽略\n\n\n\n的损失计算,避免思考能力的丢失。
demo: https://github.com/modelscope/ms-swift/blob/main/examples/train/think_model/qwen3_demo1.sh
{"messages": [{"role": "user", "content": "浙江的省会在哪?"}, {"role": "assistant", "content": "\n\n\n\n浙江的省会在杭州。"}]}
方案二:在数据集的query中额外增加/no_think,避免思考能力的丢失。
demo: https://github.com/modelscope/ms-swift/blob/main/examples/train/think_model/qwen3_demo2.sh
{"messages": [{"role": "user", "content": "浙江的省会在哪? /no_think"}, {"role": "assistant", "content": "\n\n\n\n浙江的省会在杭州。"}]}
都试了,不带思考数据,即使用非思考和掩码模式的训练,都会把思考能力丢失掉,有没有人做过类似的
Beta Was this translation helpful? Give feedback.
All reactions