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

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

True, but it looks like there’s nothing to be examined as of today. I presume the supreme goal is to train a new thinking design and after that use the very same examination metrics as o1 and the DeepSeek-R1.

Well, there must be at least some sanity check and recognition to ensure the model was trained correctly.

Oh yes, if you are speaking about the assessment variety of deepseek’s design it’s coming very soon!

As mentioned in the blog site post there is no design called Open-R1 to evaluate at all … not yet anyhow. This is a blog laying out that Hugging face will take the R1 Deepseek design, exercise how it was constructed as outlined in the paper and from what they released, and after that replicate that process.

in reality this is quite much how science works … A comes up with a strategy, discovery or development 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 currently done so … Its a blog outlining an intent to begin training a model like R1 and calling it Open-R1.

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

So there are no standards for a design that has not been constructed 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 may join myself), some prelim work done, and a strategy of attack. An excellent beginning position.

n
@edbeeching
has actually examined the launched models already

( 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 post is an introduction to the project, not a claim that we’ve recreated R1 yet. We will completely share the missing piece when we have them, you can expect the models 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 understand this remarkable 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 cost.

We will!

Excalidraw Hi n
@bojan2501
thanks, we will certainly be striving to ensure this training recipe can work for small language models on customer hardware because not everybody has a cluster of H100s in the house:-RRB- The tool we used for the images was Excalidraw! https://excalidraw.com

eagerly anticipating it! WTF are your speaking about?

should be a joke

It’s truly 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 actually never launched code or datasets of their LLM training, so I would not expect this time to be various. If they would launch it that would be amazing of course!

Yes obviously!

So generally you’re asking to change existing censorship with another flavour of censorship?

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 dealing with a paper concentrated on reproducing certain elements of DeepSeek R1. Our goal is to replicate the cold start and supply your group with a dataset that includes COT and other methods to support these efforts. We like to contribute our work to assist. Please let me understand if you find 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, but it appears like there’s nothing to be examined since today. I assume the ultimate goal is to train a new reasoning design and then utilize the same assessment metrics as o1 and the DeepSeek-R1.

That’s quite intriguing, 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 remarkable however at the very same time I wonder why they would not put these missing pieces on if they are supposed to be completely open.
Why even without reproduction and comprehension of the development they could impact a lot the market in this method?

4 replies

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

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

2 replies

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

looking forward to it! So racist articel

2 replies

WTF are your discussing?

Awesome to have this open recreation began!

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 actually cool to see how the entire open source neighborhood comes together!

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

2 replies

Ops …

Has anyone asked the DeepSeek team to publish their training information and code, or a minimum of share them privately with an independent replication project like this? Have they rejected such a demand?

A faithful duplication depends upon utilizing the very same dataset and hyperparameters. Otherwise, any major discrepancies with the published criteria would be hard to pin down-whether due to training information distinctions or the replication technique itself.

1 reply

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

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

You provide great duplication procedure of Deepseek thinking training. I will attempt something comparable to it.

This is really great information, can we tweak with specific use case when code is released?

1 reply

Yes obviously!

Please think about removing biased, or unaligned training data and make an effort to remove copyrighted works from the crawl from intake. This will make the design more functional. If you reused anthropic curation checks, this might also help, eliminate obviouslybiased data will likely add a great deal of value. We do not desire another polluted, unaligned open source model, right? And no corporate would ever use deepseek or a model that reuses it, right?
We appreciate your work for the advantage of mankind, we hope.
Miike C from NJ

1 reply

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

Can’t wait! Hopefully the model will be uncensored but whatever you can do is alright! Love seeing open source structure itself up. I’m not smart sufficient to really assist but I can contribute support lol

Hello guys, I am even just searching for code for DeepSeek-V2, in order to fully understand multi-head hidden 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 effectively described in their paper, so it would be very important to have code for this.