How to contribute

[lewtun]

Hello there 👋!

Replicating all parts of DeepSeek's R1 pipeline is going to take a community effort, especially with dataset curation and creation. If you would like to contribute, please explore the issues linked below.

[russellballestrini]

Consider this to have discussion without people creating issues.

https://docs.github.com/en/discussions/quickstart#enabling-github-discussions-on-your-organization

[qgallouedec]

Consider this to have discussion without people creating issues.

https://docs.github.com/en/discussions/quickstart#enabling-github-discussions-on-your-organization

Thanks for the suggestion. Discussions are now open #36

[pepinu]

Linking for visibility: #58

[yawar2050]

Consider this to have discussion without people creating issues.

https://docs.github.com/en/discussions/quickstart#enabling-github-discussions-on-your-organization

This really makes it easier

[Walkier]

Except for helping with dataset and documentation, how can people without access to giant clusters help contribute?

[StEvUgnIn]

Do you mind taking a look at this synthetic curated dataset with artificial knowledge from GPT-3.5?

https://docs.google.com/spreadsheets/d/1iZCuSDTpcH0qJqIOl-34tN7YXjfYfK_CbcLAkEyMJhA/edit?usp=sharing

This could probably help you figure out which data you need to retrieve in order to train R1 and R1-Zero.

[StEvUgnIn]

Could you perhaps explain or track where is Open-R1 currently with steps 1, 2 and 3?

Which tech stack and platform are you planning to use to conduct step 1 (SFT distillation)?

[elricwan]

Hello all, thanks for the great contributions, heads off.

Now I want to add few points, first, I have seen tutorials that show the training process through google colab, can we consider add few google colab examples to this repo as well for better explanation? Here is the sample:

https://www.youtube.com/watch?v=nBQ-vP5X2c0&list=WL&index=1
https://colab.research.google.com/github/unslothai/notebooks/blob/main/nb/Llama3.1_(8B)-GRPO.ipynb#scrollTo=oo4P3DhPuNPf

[StEvUgnIn]

Hello all, thanks for the great contributions, heads off.

Now I want to add few points, first, I have seen tutorials that show the training process through google colab, can we consider add few google colab examples to this repo as well for better explanation? Here is the sample:

https://www.youtube.com/watch?v=nBQ-vP5X2c0&list=WL&index=1
https://colab.research.google.com/github/unslothai/notebooks/blob/main/nb/Llama3.1_(8B)-GRPO.ipynb#scrollTo=oo4P3DhPuNPf

This notebook is interesting. The solution uses the existing R1 model. Open-R1 is a community effort to replicate the training pipeline to obtain DeepSeek-R1. Such replication will require tremendous effort and expertise compared to what is provided in this notebook.

Unsloth is a solution for quantization. Open-R1 is dedicated to pre-train R1-Zero and R1, and distill R1.

Thank you for sharing. I will watch the shared video later.

[csiefer2]

In light of DeepSeek open sourcing bits of their toolchain (https://github.com/deepseek-ai/open-infra-index) , how does this change the Open-R1 effort?

[smellslikeml]

May be out of scope for replicating Deepseek's R1, but multimodal thinking models are the new frontier, linking this discussion on making synthetic data for reasoning VLMs

https://huggingface.co/spaces/open-r1/README/discussions/10

For those interested, check out the open-source data synthesis pipeline for multimodal reasoning datasets: https://github.com/remyxai/VQASynth