Boom! Hugging Face has completely open-sourced the full training pipeline of DeepSeek-R1 and reproduced it step by step. With 26.4k stars on GitHub, this comes from the official Hugging Face team—the most authoritative take yet. DeepSeek-R1 stunned everyone, but the true training recipe remained out of reach for outsiders. Open-R1 openly reproduces the entire pipeline—from data and training to evaluation—step by step. You can follow along to understand exactly how a reasoning model is trained; it’s a textbook for anyone wanting to train or study reasoning models. For those working on large models, this is a public assignment that tears open the closed-source black box and lays it all on the table. As model capabilities grow stronger, open-sourcing the training process is what truly drives the whole industry forward.
GitHub: —

huggingface/open-r1

Source: https://github.com/huggingface/open-r1

Open R1

A fully open reproduction of DeepSeek-R1. This repo is a work in progress, let’s build it together!

Table of Contents

1. Overview
2. Plan of attack
3. Installation
4. Training models
   • SFT
   • GRPO
5. Evaluating models
6. Reproducing Deepseek’s evaluation results
7. Data generation
   • Generate data from a smol distilled R1 model
   • Generate data from DeepSeek-R1
8. Contributing

Overview

The goal of this repo is to build the missing pieces of the R1 pipeline such that everybody can reproduce and build on top of it. The project is simple by design and mostly consists of:

• src/open_r1: contains the scripts to train models as well as generate synthetic data:
  • grpo.py: trains a model with GRPO on a given dataset.
  • sft.py: performs a simple SFT of a model on a dataset.
  • generate.py: generates synthetic data from a model using Distilabel (https://github.com/argilla-io/distilabel).
• Makefile: contains easy-to-run commands for each step in the R1 pipeline leveraging the scripts above.

Plan of attack

We will use the DeepSeek-R1 tech report (https://github.com/deepseek-ai/DeepSeek-R1) as a guide, which can roughly be broken down into three main steps:

• Step 1: replicate the R1-Distill models by distilling a high-quality corpus from DeepSeek-R1.
• Step 2: replicate the pure RL pipeline that DeepSeek used to create R1-Zero. This will likely involve curating new, large-scale datasets for math, reasoning, and code.
• Step 3: show we can go from base model to RL-tuned via multi-stage training.

News 🗞️

• 🧑🍳 [2025/05/26] (Step 1 completed!) We release Mixture-of-Thoughts (https://huggingface.co/datasets/open-r1/Mixture-of-Thoughts)–a curated reasoning dataset of 350k verified traces distilled from R1. The dataset spans tasks in mathematics, coding, and science, and is designed to teach language models to reason step-by-step. We also provide a recipe to train OpenR1-Distill-7B (https://huggingface.co/open-r1/OpenR1-Distill-7B), which replicates the reasoning capabilities of deepseek-ai/DeepSeek-R1-Distill-Qwen-7B (https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-7B) and marks the completion of step 1 in the Open R1 project.
• ⚡️ [2025/03/11] (update #3) (https://huggingface.co/blog/open-r1/update-3): We release the CodeForces-CoTs (https://huggingface.co/datasets/open-r1/codeforces-cots) dataset of 10k competitive programming problems and 100k solutions distilled from R1. We also release IOI24: a new benchmark of very hard problems from international olympiads. A 7B Qwen model trained on CodeForces-CoTs can outperform Claude 3.7 Sonnet on IOI24, while a 32B model can outperform R1 itself.
• ∞ [2025/02/10] (update #2) (https://huggingface.co/blog/open-r1/update-2): We release the OpenR1-Math-220k (https://huggingface.co/datasets/open-r1/OpenR1-Math-220k) dataset of 220k traces distilled from R1 on a new version of NuminaMath. Models trained on this dataset match the performance of DeepSeek’s distilled ones.
• 🔥 [2025/02/02] (update #1) (https://huggingface.co/blog/open-r1/update-1): We implement the first parts of the training (https://github.com/huggingface/open-r1?tab=readme-ov-file#training-models), inference (https://github.com/huggingface/open-r1?tab=readme-ov-file#data-generation), and evaluation (https://github.com/huggingface/open-r1?tab=readme-ov-file#reproducing-deepseeks-evaluation-results) pipelines. Let’s go!

Installation

Libraries rely on CUDA 12.4. If you see errors related to segmentation faults, double check the version your system is running with nvcc --version.

To run the code in this project, first, create a Python virtual environment using e.g. uv. To install uv, follow the UV Installation Guide (https://docs.astral.sh/uv/getting-started/installation/).

As a shortcut, run make install to setup development libraries (spelled out below). Afterwards, if everything is setup correctly you can try out the Open-R1 models.

shell uv venv openr1 --python 3.11 && source openr1/bin/activate && uv pip install --upgrade pip

For Hugging Face cluster users, add export UV_LINK_MODE=copy to your .bashrc to suppress cache warnings from uv

Next, install vLLM and FlashAttention:

shell uv pip install vllm==0.8.5.post1 uv pip install setuptools && uv pip install flash-attn --no-build-isolation

This will also install PyTorch v2.6.0 and it is very important to use this version since the vLLM binaries are compiled for it.

You can then install the remaining dependencies for your specific use case via pip install -e .[LIST OF MODES]. For most contributors, we recommend:

shell GIT_LFS_SKIP_SMUDGE=1 uv pip install -e ".[dev]"

Next, log into your Hugging Face and Weights and Biases accounts as follows:

shell huggingface-cli login wandb login

Finally, check whether your system has Git LFS installed so that you can load and push models/datasets to the Hugging Face Hub:

shell git-lfs --version

If it isn’t installed, run:

shell sudo apt-get install git-lfs

Training models

The training commands below are configured for a node of 8 x H100s (80GB). For different hardware and topologies, you may need to tune the batch size and number of gradient accumulation steps.

We support training models with either DDP or DeepSpeed (ZeRO-2 and ZeRO-3). For example, to perform SFT on a dataset distilled from DeepSeek-R1 with reasoning traces such as open-r1/Mixture-of-Thoughts (https://huggingface.co/datasets/open-r1/Mixture-of-Thoughts), run:

``shell

Train via command line

accelerate launch –config_file=recipes/accelerate_configs/zero3.yaml src/open_r1/sft.py \
–model_name_or_path open-r1/Qwen2.5-Math-7B-RoPE-300k \
–dataset_name open-r1/Mixture-of-Thoughts \
–dataset_config all \
–eos_token ‘<|im_end|>’ \
–learning_rate 4.0e-5 \
–num_train_epochs 5 \
–max_seq_length 32768 \
–per_device_train_batch_size 2 \
–gradient_checkpointing \
–bf16 \
–use_liger_kernel \
–output_dir data/OpenR1-Distill-7B

Train via YAML config

accelerate launch –config_file recipes/accelerate_configs/zero3.yaml src/open_r1/sft.py \
–config recipes/OpenR1-Distill-7B/sft/config_distill.yaml
``

Currently, the following tasks are supported:

• Supervised Fine-Tuning sft
• Group Relative Policy Optimization grpo

If you scale up/down the number of GPUs, we recommend also scaling up the per-device batch size or number of gradient accumulation steps to keep the global batch size constant.

By default, these scripts will push each model to your Hugging Face Hub username, i.e. {username}/{model_name}-{task}. You can override the parameters in each YAML config by appending them to the command as follows:

``shell

Change the base model to a smaller variant

accelerate launch –config_file recipes/accelerate_configs/zero3.yaml src/open_r1/sft.py \
–config recipes/OpenR1-Distill-7B/sft/config_distill.yaml \
–model_name_or_path Qwen/Qwen3-0.6B-Base \
–hub_model_id OpenR1-Distill-0.6B \
–output_dir data/OpenR1-Distill-0.6B
``

If you also wish to override the Weights and Biases default settings, you can do so as follows:

shell accelerate launch --config_file recipes/accelerate_configs/zero3.yaml src/open_r1/sft.py \ --config recipes/OpenR1-Distill-7B/sft/config_distill.yaml --wandb_entity huggingface --wandb_project open-r1 --run_name Qwen2.5-1.5B-GRPO

🚨 WARNING 🚨
Most base models like meta-llama/Llama-3.2-1B do not have a chat template, so we set ChatML as the default during training. However, for Qwen base models like Qwen/Qwen2.5-1.5B, a chat template is pre-defined in the tokenizer, so the EOS token must be set accordingly, e.g.

``diff

Align EOS token with chat template for Qwen base models

accelerate launch –config_file=recipes/accelerate_configs/zero3.yaml src/open_r1/sft.py \
–model_name_or_path Qwen/Qwen2.5-1.5B \

• –eos_token ‘<|im_end|>’
  –dataset_name open-r1/Mixture-of-Thoughts \
  –dataset_config all \
  –learning_rate 4.0e-5 \
  –num_train_epochs 1 \
  –max_seq_length 32768 \
  –per_device_train_batch_size 16 \
  –gradient_checkpointing \
  –bf16 \
  –use_liger_kernel \
  –output_dir data/Qwen2.5-1.5B-Open-R1-Distill
  ``

If you wish to use a custom chat template (e.g. Llama or Gemma), then the chat template and associated EOS token must be provided:

``diff

Align EOS token with custom chat template

accelerate launch –config_file=recipes/accelerate_configs/zero3.yaml src/open_r1/sft.py \
–model_name_or_path meta-llama/Llama-3.2-1B \

• –chat_template “$(cat llama_chat_template.jinja)” \
• –eos_token ‘<|eot_id|>’ \
  –dataset_name open-r1/Mixture-of-Thoughts \
  –dataset_config all \
  –learning_rate 4.0e-5 \
  –num_train_epochs 1 \
  –max_seq_length 32768 \
  –per_device_train_batch_size 16 \
  –gradient_checkpointing \
  –bf16 \
  –use_liger_kernel \
  –output_dir data/Llama-3.2-1B-Open-R1-Distill
  ``

SFT distillation

We provide a recipe to reproduce the reasoning capabilities of deepseek-ai/DeepSeek-R1-Distill-Qwen-7B (https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-7B), starting from the same base model. To do so, run:

shell ACCELERATE_LOG_LEVEL=info accelerate launch --config_file recipes/accelerate_configs/zero3.yaml \ src/open_r1/sft.py \ --config recipes/OpenR1-Distill-7B/sft/config_distill.yaml

The result will be a model like open-r1/OpenR1-Distill-7B (https://huggingface.co/open-r1/OpenR1-Distill-7B), with the following downstream performance:

You can adjust the YAML config to train on a different base model or dataset.

GRPO

We use TRL’s vLLM backend (https://huggingface.co/docs/trl/speeding_up_training?vllm+examples=GRPO#vllm-for-fast-generation-in-online-methods) to scale training to large models across multiple nodes. For single-node training of smol models across 8 GPUs, use vllm_mode="colocate" to run vLLM in the same process as the training script:

shell ACCELERATE_LOG_LEVEL=info \ accelerate launch --config_file recipes/accelerate_configs/zero3.yaml \ src/open_r1/grpo.py --config recipes/DeepSeek-R1-Distill-Qwen-1.5B/grpo/config_demo.yaml \ --vllm_mode colocate

The chat template used in the distilled DeepSeek models omits the contents of the reasoning block within the and tags. It also prefills the assistant response with `` which interferes with the format reward function. To handle that, it is important to override the chat template as done in e.g. recipes/DeepSeek-R1-Distill-Qwen-1.5B/grpo/config_demo.yaml.

For multi-node training on N+1 nodes, with 1 node running the vLLM server and N nodes running training, we provide an example Slurm script. For example, to run the above example on 1+1 nodes with data parallelism, run:

shell sbatch --nodes=2 slurm/train.slurm --model Qwen2.5-1.5B-Instruct --task grpo --config demo --accelerator zero2 --dp 8 --tp 1

See the Launching jobs on a Slurm cluster section for more details.

GRPO dataset filtering

We provide support to filter datasets by generating and computing pass rate on veriable tasks, see this README

👨💻 Training with a code interpreter

We provide a code reward function for executing code generated by the policy during training. Currently, this reward function targets code contests like Codeforces (https://codeforces.com), where solutions are executed against a set of test cases and the overall success rate is returned as the final reward. To ensure safe execution, we support multiple sandbox providers:

1. E2B (https://e2b.dev) - Fast, cloud-based sandboxes with focus on Python execution
2. Morph (https://cloud.morph.so/web/) - Cloud-based sandboxes with broader language support - Python/JS/C++/Rust

To use the code reward function, first install the necessary dependencies:

shell uv pip install -e '.[code]'

E2B Provider

To use E2B sandboxes, create a .env file and add your E2B API token:

E2B_API_KEY="e2b_xxx"

Morph Provider

To use Morph, first install the morphcloud package:

shell pip install morphcloud

Then add your Morph API token to the .env file:

MORPH_API_KEY="YOUR_MORPH_API_KEY"

To specify which provider to use, add the provider_type parameter in your configuration:

``yaml

For E2B

provider_type: e2b

For Morph

provider_type: morph
``

Dataset Requirements

Make sure your dataset contains a verification_info column with the following schema (adopted from PrimeIntellect’s excellent datasets (https://huggingface.co/collections/PrimeIntellect/synthetic-1-67a2c399cfdd6c9f7fae0c37) of verifiable problems):

python { "language": "python", # Morph supports more languages including C++, Java, etc. "test_cases": [ { "input": "4\n4\n0001\n1000\n0011\n0111\n3\n010\n101\n0\n2\n00000\n00001\n4\n01\n001\n0001\n00001\n", "output": "1\n3 \n-1\n0\n\n2\n1 2 \n", "type": "stdin_stdout", } ], }

For example, to train a smol model on Python problems, start the vLLM server:

shell CUDA_VISIBLE_DEVICES=0 trl vllm-serve --model Qwen/Qwen2.5-1.5B-Instruct

Then run training with:

shell CUDA_VISIBLE_DEVICES=1,2,3,4,5,6,7 ACCELERATE_LOG_LEVEL=info \ accelerate launch --config_file recipes/accelerate_configs/zero2.yaml --num_processes=7 \ src/open_r1/grpo.py --config recipes/Qwen2.5-1.5B-Instruct/grpo/config_demo_code.yaml

Using Router Services

It is possible to be rate limited when too many scripts are executed on sandbox services. For both providers, we offer router scripts that can be launched on a CPU node:

For E2B:
shell sbatch slurm/e2b_router.slurm

For Morph:
shell sbatch slurm/morph_router.slurm

Then add the router URL in your training YAML config:

``yaml

For E2B

e2b_router_url: 1.2.3.4:8000

For Morph

morph_router_url: 1.2.3.4:8000
``

The port should match the one used when launching the router. All training jobs can share the same router IP which will ensure parallel executions are properly managed.

Competitive Programming problems: IOI & CodeForces

We provide ioi_code_reward and cf_code_reward reward functions for executing problems from IOI (https://hf.co/datasets/open-r1/ioi) and CodeForces (https://huggingface.co/datasets/open-r1/codeforces), respectively. You can use either piston (https://github.com/engineer-man/piston) or Morph (currently IOI only) as your execution provider.

Piston

To use Piston:

1. Get piston workers running, see slurm/piston/README.md
2. Set your environment variable PISTON_ENDPOINTS to slurm or to a list of piston worker endpoints

For IOI:
3. In your configuration, use ioi_provider: "piston"

For CodeForces:
3. Download the generated (hard) test cases:
``

change PATH_TO_SAVE_TESTCASES. Increase –max-workers according to your machine’s capacity

huggingface-cli download open-r1/codeforces –repo-type=dataset –include=‘generated_tests/*.parquet’ –max-workers=8 –local-dir PATH_TO_SAVE_TESTCASES
4. Save the path in .env:
CF_TESTS_FOLDER=PATH_TO_SAVE