Open-R1: a Fully Open Reproduction Of DeepSeek-R1

Hey there! This post is an intro to the task, not a claim that we’ve recreated R1 yet. We’re integrating in the open, so as soon as we have evaluation numbers, we’ll share them. You can follow our development on Hugging Face and GitHub.

True, however it looks like there’s nothing to be assessed as of today. I presume the supreme goal is to train a brand-new thinking design and then use the exact same assessment metrics as o1 and the DeepSeek-R1.

Well, there need to be at least some sanity check and validation to make sure the model was trained correctly.

Oh yes, if you are speaking about the evaluation variety of deepseek’s model it’s coming soon!

As discussed in the blog site post there is no model called Open-R1 to evaluate at all … not yet anyway. This is a blog site describing that Hugging face will take the R1 Deepseek model, work out how it was built as laid out in the paper and from what they launched, and then replicate that process.

in fact this is basically how science works … A creates a plan, discovery or innovation and it is evaluated by B, C and D to see if it is reproduceable. Thats been the foundation of research now for a couple of centuries.

This blog site is not saying they have actually already done so … Its a blog site outlining an intent to start training a model like R1 and calling it Open-R1.

Also DeepSeek-R1 was only launched recently, and even in their paper they outlined the compute hours required. While those are low compute hours for a SOTA model this does not imply you can train said model in a week. I ‘d personally love to be able to train a transformer model in a week, however we might need to wait a while for that level of calculate technology.

So there are no criteria for a model that has not been developed yet right? As described in the blog, and once again in reply to your concern.

However fear not, there is a GitHub Repo already and contributors (hell I might join myself), some prelim work done, and a strategy of attack. A good starting position.

n
@edbeeching
has assessed the launched models currently

( src: https://x.com/edwardbeeching/status/1884273209136275742)

R1 just trained on o1 outputs, so jointly …/ s. This is what the new AI czars are saying

Hi! This article is an intro to the project, not a claim that we have actually replicated R1 yet. We will completely share the missing out on piece when we have them, you can expect the designs and datasets to be upload in this Hugging Face org and the code to be in this GitHub repo

That’s great and important to comprehend this incredible buzz that does not have technical understanding and explanation. Science has to do with reproduction, and if they claim to be open, let them fullfill the open part.

Please do publish the training expense.

We will!

Excalidraw Hi n
@bojan2501
thanks, we will indeed be working hard to make certain this training recipe can work for little language models on consumer hardware given that not everyone has a cluster of H100s in your home:-RRB- The tool we utilized for the images was Excalidraw! https://excalidraw.com

eagerly anticipating it! WTF are your talking about?

should be a joke

It’s really cool to see how the entire open source community comes together!

Ops …

5.5 M is number reporter in the deepseekv3 tech report (simply the training, not the experiment afaik), for R1 hard to estimate tbh however much less than 5.5 M imo

Historically, they have never ever launched code or datasets of their LLM training, so I would not anticipate this time to be various. If they would release it that would be incredible obviously!

Yes of course!

So essentially you’re asking to replace existing censorship with another flavour of ?

The code for the designs are inside the model repositories, e.g. for V3: https://huggingface.co/deepseek-ai/DeepSeek-V3/blob/main/modeling_deepseek.py

Hello Team, I’m Ray Bernard, the author and creator of EQUATOR. My research team will be working on a paper focused on duplicating specific components of DeepSeek R1. Our goal is to reproduce the cold start and offer your group with a dataset that includes COT and other strategies to support these efforts. We like to contribute our work to assist. Please let me know if you discover this beneficial. Best, Ray Bernard https://www.facebook.com/groups/1186310571520299/

Where is the evaluation numbers? without it you can’t call it recreation.

8 replies

True, however it looks like there’s nothing to be evaluated as of right now. I assume the ultimate objective is to train a new reasoning model and after that use the very same assessment metrics as o1 and the DeepSeek-R1.

That’s rather interesting, I was asking myself why the concerns the author exposed here are not being asked by others? I think the work they have actually done is unforgettable however at the very same time I wonder why they would not put these missing out on pieces on if they are supposed to be totally open.
Why even without recreation and understanding of the development they could impact a lot the marketplace in this method?

4 replies

Hi! This blog site post is an intro to the project, not a claim that we’ve recreated R1 yet. We will absolutely share the missing piece when we have them, you can anticipate the designs and datasets to be upload in this Hugging Face org and the code to be in this GitHub repo

Interesting read, and it is good that we see more effort into this instructions: more optimization and less strength.
Also question what tool did the author use for developing action diagram.

2 replies

Excalidraw I’m so glad that effort like this already exist, I’m gon na attempt to contribute:-RRB- 1 reply

anticipating it! So racist articel

2 replies

WTF are your speaking about?

Awesome to have this open recreation started!

For Step # 1 check out https://github.com/open-thoughts/open-thoughts!

https://x.com/ryanmart3n/status/1884284101265612856

Let’s do this thing!

1 reply

It’s really cool to see how the entire open source community comes together!

Does anyone know the actual training expense of r1? I can’t discover it in the paper or the statement post. Is the 6M cost reported by media simply the number taken from v3’s training cost?

2 replies

Ops …

Has anyone asked the DeepSeek group to publish their training data and code, or at least share them privately with an independent duplication task like this? Have they turned down such a request?

A devoted replication depends upon using the very same dataset and hyperparameters. Otherwise, any significant inconsistencies with the released criteria would be tough to pin down-whether due to training information differences or the duplication method itself.

1 reply

Historically, they have actually never released code or datasets of their LLM training, so I would not expect this time to be different. If they would release it that would be amazing obviously!

In the meantime we need to make finest guess price quotes and see if we can get there ourselves.

You provide great replication procedure of Deepseek thinking training. I will try something similar to it.

This is actually great information, can we tweak with specific use case when code is launched?

1 reply

Yes of course!

Please consider eliminating biased, tainted or unaligned training information and make an effort to get rid of copyrighted works from the crawl from intake. This will make the model more usable. If you recycled anthropic curation checks, this may also assist, remove obviouslybiased information will likely add a great deal of worth. We do not want another polluted, unaligned open source design, right? And no corporate would ever utilize deepseek or a design that reuses it, right?
We appreciate your work for the benefit of humankind, we hope.
Miike C from NJ

1 reply

So basically you’re asking to replace existing censorship with another flavour of censorship?

Can’t wait! Hopefully the design will be uncensored however whatever you can do is alright! Love seeing open source structure itself up. I’m not smart adequate to in fact assist however I can contribute moral support lol

Hello guys, I am even just trying to discover code for DeepSeek-V2, in order to completely understand multi-head latent attention. You do not seem to have code in Hugging Face even for that. Or am I missing out on something? Don’t see anything in src/transformers/models. MLA is not appropriately described in their paper, so it would be essential to have code for this.