Neura Market
open-r1 logo

open-r1

Free

Fully open reproduction of DeepSeek-R1

FreeFree tier
Type
Open Source
Company
Hugging Face
Visit Website GitHub (38 stars)

About open-r1

Open R1 is an open-source project by Hugging Face aimed at fully reproducing DeepSeek-R1, an advanced reasoning model. The repository provides scripts for training models using GRPO (Group Relative Policy Optimization) and SFT (Supervised Fine-Tuning), as well as generating synthetic data with Distilabel. The project follows a three-step plan: distilling a high-quality corpus from DeepSeek-R1 to create R1-Distill models, replicating the pure RL pipeline for R1-Zero using curated datasets for math, reasoning, and code, and demonstrating multi-stage training from a base model to RL-tuned. As of May 2025, Step 1 has been completed with the release of the Mixture-of-Thoughts dataset (350k verified reasoning traces) and a recipe for OpenR1-Distill-7B, which matches DeepSeek’s distilled model performance. Additional released datasets include CodeForces-CoTs and OpenR1-Math-220k. The project is community-driven and a work in progress, requiring CUDA 12.4 and Python with uv for setup.

Key Features

Fully open-source reproduction pipeline for DeepSeek-R1
Training scripts: GRPO (grpo.py), SFT (sft.py), and synthetic data generation (generate.py)
Published high-quality reasoning datasets: Mixture-of-Thoughts (350k traces), CodeForces-CoTs, OpenR1-Math-220k
Recipe for OpenR1-Distill-7B that matches DeepSeek-R1-Distill-Qwen-7B performance
Three-step plan: distillation, pure RL (R1-Zero), and multi-stage training
Easy setup with Makefile and Python virtual environment via uv

Pros & Cons

Pros
  • Completely open-source, allowing full transparency and community contributions
  • Releases verified training datasets and model recipes to accelerate research
  • Actively maintained with frequent updates and clear roadmap
  • Leverages Hugging Face ecosystem and existing tools like transformers and accelerate
  • Provides reproducible training scripts (GRPO, SFT) with minimal dependencies
Cons
  • Work in progress – not all steps of DeepSeek-R1 replication are completed yet
  • Requires CUDA 12.4 and specific GPU hardware for training
  • Dependency on proprietary DeepSeek-R1 model for distillation data generation
  • Limited documentation and setup instructions for non-expert users

Best For

Reproducing and studying reasoning capabilities of DeepSeek-R1Training custom reasoning models via distillation or reinforcement learningGenerating synthetic reasoning data for math, code, and science tasksBenchmark evaluation on math and coding problems (e.g., IOI24)

FAQ

What is Open R1?
Open R1 is a Hugging Face project to fully reproduce DeepSeek-R1 as an open-source implementation, providing training scripts, datasets, and recipes.
Is Open R1 fully functional?
Not yet. The project is a work in progress. Step 1 (distillation) is completed, but Steps 2 and 3 (pure RL and multi-stage training) are still under development.
What datasets are available?
Open R1 has released Mixture-of-Thoughts (350k reasoning traces), CodeForces-CoTs (100k solutions), and OpenR1-Math-220k. These are distilled from DeepSeek-R1.
What hardware is required?
Training requires GPUs with CUDA 12.4. The project recommends using Python and uv for environment setup.