[GH-ISSUE #320] 请问R1训练是不是只做了论文中写的2.3.3和2.3.4的内容,而2.3.1和2.3.2实际上只是为了生成数据给2.3.3用的? #249

Closed
opened 2026-10-26 21:17:51 +00:00 by bsakaguchi · 3 comments
Owner

Originally created by @ciaoyizhen on GitHub (Feb 8, 2025).
Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-R1/issues/320

跟朋友在聊这个论文
网上很多图是长这样的
Image

这个图是不是错的?

论文里2.3.3里面写到
We fine-tune DeepSeek-V3-Base for two epochs using the above curated dataset of about 800k samples.
然后摘要里也写到

and then retrain the DeepSeek-V3-Base model.

所以是不是论文中写的sft-rl-sft-rl的这四个阶段,实际上前两个阶段只是训了一个模型,然后用这个模型生成了数据,然后又是从V3-base开始训一个sft-rl就好了

想知道 是不是训练实际上也还是sft+rl?

求大佬回复一下!!!谢谢🙏

Originally created by @ciaoyizhen on GitHub (Feb 8, 2025). Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-R1/issues/320 跟朋友在聊这个论文 网上很多图是长这样的 ![Image](https://github.com/user-attachments/assets/8402ffcc-cb14-4fa4-b305-9f60db40cf29) 这个图是不是错的? 论文里2.3.3里面写到 `We fine-tune DeepSeek-V3-Base for two epochs using the above curated dataset of about 800k samples.` 然后摘要里也写到 ``` and then retrain the DeepSeek-V3-Base model. ``` 所以是不是论文中写的sft-rl-sft-rl的这四个阶段,实际上前两个阶段只是训了一个模型,然后用这个模型生成了数据,然后又是从V3-base开始训一个sft-rl就好了 想知道 是不是训练实际上也还是sft+rl? 求大佬回复一下!!!谢谢🙏
Author
Owner

@zrmzack commented on GitHub (Feb 8, 2025):

去看下youtube 解说r1 和v3 的论文,详细描述了这个v3---r1; r1只有rl了

<!-- gh-comment-id:2644447559 --> @zrmzack commented on GitHub (Feb 8, 2025): 去看下youtube 解说r1 和v3 的论文,详细描述了这个v3---r1; r1只有rl了
Author
Owner

@sflowing commented on GitHub (Feb 13, 2025):

Good question, and I have the exact same question!

Image

Is this pipeline right?

<!-- gh-comment-id:2656065791 --> @sflowing commented on GitHub (Feb 13, 2025): Good question, and I have the exact same question! ![Image](https://github.com/user-attachments/assets/57483b15-644f-4c2b-88c8-77ba500a21e7) Is this pipeline right?
Author
Owner

@ciaoyizhen commented on GitHub (Feb 19, 2025):

@sflowing
video of huggingface about deepseek

I think you are right.

<!-- gh-comment-id:2667818336 --> @ciaoyizhen commented on GitHub (Feb 19, 2025): @sflowing [video of huggingface about deepseek](https://www.youtube.com/watch?v=1xDVbu-WaFo) I think you are right.
Sign in to join this conversation.
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
deepseek-ai/DeepSeek-R1#249
No description provided.