Welcome to Oumi 🙌

Oumi Friend,

What a time it’s been since stepping out of stealth at the end of January…

🚀 Oumi Launch Recap & Highlights:

• ⭐ 7.6K+ GitHub Stars

• 🌐 900+ Discord Community Members

• 🔬 120+ research participants

• And this is just the beginning… 

We’re thrilled to bring you the first of what will be a regular newsletter from the Oumi team. This will serve as your insider’s guide to everything from developments (research papers, model release, etc.) to community highlights you might have missed. We’ll have a surprise or special edition every once in a while too, to keep things even more interesting.

✋ Get Involved

✍️ Start Contributing:

Whether you’re a seasoned AI researcher or just dipping your toes into open-source, there’s a GitHub issue with your name on it. Browse open issues here.

💡 Tip: You can filter for issues with the "good first issue" tag to get started!

📣 Spread the Word:

If you haven’t already, give our GitHub repository a star ⭐(and maybe ask a friend to do the same!). Every bit of support helps us grow the community and accelerate open-source AI innovation.

✋ Join a Research Project:

Browse proposed ideas you may be interested in, or propose a new idea entirely here. Collaboration is what Oumi’s all about.

👀 Releases you might have missed:

MiniMath-R1:

We collected 650,000+ R1 prompts and responses for you to train your own DeepSeek-R1 model. With this data we made MiniMath-R1-1.5B, the top MMLU-Pro-Math model at <=1.5B parameters, achieving an accuracy of 44.4% vs 43.0% from Qwen2.5-1.5B, and 38.5% from the R1 1.5B distill, +1.4 over the SOTA and +6 over the base model.

Interested in training your own DeepSeek AI R1 model? Here’s the data, model, and notebook where you can reproduce our results:

• 📈 Datasets: 

  • MetaMathQA-R1 

  • lmsys_chat_1m_clean_R1  

• 💻 Model

• 📓 Notebook

GRPO:

We now support GRPO training in Oumi! This is a new reinforcement learning algorithm introduced by the DeepSeek team, used for training reasoning in DeepSeek-R1. Check out our example here, and stay tuned for a more in-depth tutorial coming soon!

• 📁 GRPO configuration file

QwQ 32B:

We now support QwQ training, inference, and evaluation in Oumi! This brand-new reasoning model by Alibaba, based on Qwen2.5, performs on par with DeepSeek-R1 for several math and coding tasks.

• 📁 LoRA training configuration file

• 📁 Inference configuration file

• 📁 Evaluation configuration file

🙏 Last but not least: THANK YOU!

We know your inbox (and attention) is precious. Thank you for being part of the Oumi community and for reading our newsletter. 

We’ll be back soon with more updates, new research initiatives, and (hopefully) a few surprises.  👀 

In the meantime, let’s keep shaping the future of open-source AI together!

Thanks for reading and contributing,

The Oumi Team

Have questions or feedback? Reply to this email or drop into our Discord community. We’d love to hear from you!