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 on MATH-500
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 and evaluate 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.
  • evaluate.py: evaluates a model on the R1 benchmarks.
  • generate.py: generates synthetic data from a model using 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 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.

Installation

Note: Libraries rely on CUDA 12.4. Double check your system if you get segmentation faults.

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.

uv venv openr1 --python 3.11 && source openr1/bin/activate && uv pip install --upgrade pip --link-mode=copy

Next, install vLLM:

uv pip install vllm==0.7.1 --link-mode=copy

This will also install PyTorch v2.5.1 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:

GIT_LFS_SKIP_SMUDGE=1 uv pip install -e ".[dev]" --link-mode=copy

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

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:

git-lfs --version

If it isn't installed, run:

sudo apt-get install git-lfs

Training models

We support training models with either DDP or DeepSpeed (ZeRO-2 and ZeRO-3). For example, to run SFT on a dataset distilled from DeepSeek-R1 with reasoning traces such as Bespoke-Stratos-17k, run:

# Train via command line
accelerate launch --config_file=recipes/accelerate_configs/zero3.yaml src/open_r1/sft.py \
    --model_name_or_path Qwen/Qwen2.5-1.5B-Instruct \
    --dataset_name HuggingFaceH4/Bespoke-Stratos-17k \
    --learning_rate 2.0e-5 \
    --num_train_epochs 1 \
    --packing \
    --max_seq_length 4096 \
    --per_device_train_batch_size 2 \
    --gradient_accumulation_steps 8 \
    --gradient_checkpointing \
    --bf16 \
    --output_dir data/Qwen2.5-1.5B-Open-R1-Distill

# Train via YAML config
accelerate launch --config_file recipes/accelerate_configs/zero3.yaml src/openr1/sft.py \
    recipes/Qwen/Qwen2.5-1.5B-Instruct/sft/config_demo.yaml

Currently, the following tasks are supported:

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

Tip

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:

# Change batch size, number of epochs etc
accelerate launch --config_file recipes/accelerate_configs/zero3.yaml src/openr1/sft.py \
    recipes/Qwen/Qwen2.5-1.5B-Instruct/sft/config_demo.yaml
    --per_device_train_batch_size=1 --num_train_epochs=5

Note

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.

SFT

To run SFT on a dataset distilled from DeepSeek-R1 with reasoning traces such as Bespoke-Stratos-17k, run:

ACCELERATE_LOG_LEVEL=info accelerate launch --config_file recipes/accelerate_configs/zero3.yaml \
    src/open_r1/sft.py \
    --config recipes/Qwen2.5-1.5B-Instruct/sft/config_demo.yaml

GRPO

To train via the GRPO trainer, we use one GPU to run vLLM for faster generation and the remaining GPUs for training. For example, one a node with 8 GPUs, use the recipes/accelerate_configs/zero3.yaml config and then overwrite num_processes to run on 7 devices:

ACCELERATE_LOG_LEVEL=info accelerate launch --config_file recipes/accelerate_configs/zero3.yaml \
    --num_processes=7 src/open_r1/grpo.py \
    --config recipes/Qwen2.5-1.5B-Instruct/grpo/config_demo.yaml

We provide a minimal reproducible experiment using GRPO for mathematical reasoning, referencing the approach from SimpleRL-Reason which uses a 7B model trained on 8K examples. Running this on 8 H100 80G GPU takes about 3 hours:

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-Math-7B/grpo/config_simple_rl.yaml

Our final model, while using different learning rates, loss functions and reward structures, achieves 69.4% accuracy on MATH-500, demonstrating a 17%+ improvement over the base model.

Launching jobs on a Slurm cluster

If you have access to a Slurm cluster, we provide a slurm/train.slurm script that will automatically queue training jobs for you. Here's how you can use it:

sbatch --job-name=open_r1 --nodes=1 slurm/train.slurm {model_name} {task} {config_suffix} {accelerator}

Here {model_name} and {task} are defined as above, while {config_suffix} refers to the specific config and {accelerator} refers to the choice of 🤗 Accelerate config in recipes/accelerate_configs. If you wish to override the default config parameters, you can provide them by appending a space-separated string like '--arg1=value1 --arg2=value2'. Here's a concrete example to run SFT on 1 node of 8 GPUs:

# Launch on Slurm and override default hyperparameters
sbatch --job-name=open_r1 --nodes=1 slurm/train.slurm Qwen2.5-1.5B-Instruct sft demo zero3 '--per_device_train_batch_size=1 --num_train_epochs=5'

You can scale the number of nodes by increasing the --nodes flag.

Note

The configuration in slurm/train.slurm is optimised for the Hugging Face Compute Cluster and may require tweaking to be adapted to your own compute nodes.

Evaluating models

We use lighteval to evaluate models, with custom tasks defined in src/open_r1/evaluate.py. For models which fit on a single GPU, run:

MODEL=deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B
MODEL_ARGS="pretrained=$MODEL,dtype=bfloat16,max_model_length=32768,gpu_memory_utilisation=0.8"
OUTPUT_DIR=data/evals/$MODEL

# AIME 2024
TASK=aime24
lighteval vllm $MODEL_ARGS "custom|$TASK|0|0" \
    --custom-tasks src/open_r1/evaluate.py \
    --use-chat-template \
    --output-dir $OUTPUT_DIR

# MATH-500
TASK=math_500
lighteval vllm $MODEL_ARGS "custom|$TASK|0|0" \
    --custom-tasks src/open_r1/evaluate.py \
    --use-chat-template \
    --output-dir $OUTPUT_DIR

# GPQA Diamond
TASK=gpqa:diamond
lighteval vllm $MODEL_ARGS "custom|$TASK|0|0" \
    --custom-tasks src/open_r1/evaluate.py \
    --use-chat-template \
    --output-dir $OUTPUT_DIR 

Important

You must set max_model_length=32768 in the vllm command to align with the generation_size we define per eval. Without this, lighteval will throw an error.

To increase throughput across multiple GPUs, use data parallel as follows:

NUM_GPUS=8
MODEL=deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B
MODEL_ARGS="pretrained=$MODEL,dtype=bfloat16,data_parallel_size=$NUM_GPUS,max_model_length=32768,gpu_memory_utilisation=0.8"
TASK=aime24
OUTPUT_DIR=data/evals/$MODEL

lighteval vllm $MODEL_ARGS "custom|$TASK|0|0" \
    --custom-tasks src/open_r1/evaluate.py \
    --use-chat-template \
    --output-dir $OUTPUT_DIR 

For large models which require sharding across GPUs, use tensor parallel and run:

NUM_GPUS=8
MODEL=deepseek-ai/DeepSeek-R1-Distill-Qwen-32B
MODEL_ARGS="pretrained=$MODEL,dtype=bfloat16,tensor_parallel_size=$NUM_GPUS,max_model_length=32768,gpu_memory_utilisation=0.8"
TASK=aime24
OUTPUT_DIR=data/evals/$MODEL

export VLLM_WORKER_MULTIPROC_METHOD=spawn
lighteval vllm $MODEL_ARGS "custom|$TASK|0|0" \
    --custom-tasks src/open_r1/evaluate.py \
    --use-chat-template \
    --output-dir $OUTPUT_DIR 

You can also launch an evaluation with make evaluate, specifying the model, task, and optionally the parallelism technique and number of GPUs.

To evaluate on a single GPU:

make evaluate MODEL=deepseek-ai/DeepSeek-R1-Distill-Qwen-32B TASK=aime24

To use Data Parallelism:

make evaluate MODEL=deepseek-ai/DeepSeek-R1-Distill-Qwen-32B TASK=aime24 PARALLEL=data NUM_GPUS=8

To use Tensor Parallelism:

make evaluate MODEL=deepseek-ai/DeepSeek-R1-Distill-Qwen-32B TASK=aime24 PARALLEL=tensor NUM_GPUS=8

Reproducing Deepseek's evaluation results

Note

The DeepSeek-R1 paper uses sampling with a temperature of 0.6, a top-p value of 0.95, and 64 responses per query to estimate pass@1. Below, we report the results from greedy decoding, which likely explains the small 1-3σ discrepancies between our results and theirs.

MATH-500

We are able to reproduce Deepseek's reported results on the MATH-500 benchmark within ~1-3 standard deviations:

Model	MATH-500 (🤗 LightEval)	MATH-500 (DeepSeek Reported)
DeepSeek-R1-Distill-Qwen-1.5B	81.2	83.9
DeepSeek-R1-Distill-Qwen-7B	91.8	92.8
DeepSeek-R1-Distill-Qwen-14B	94.2	93.9
DeepSeek-R1-Distill-Qwen-32B	95.0	94.3
DeepSeek-R1-Distill-Llama-8B	85.4	89.1
DeepSeek-R1-Distill-Llama-70B	93.4	94.5

To reproduce these results use the following command:

NUM_GPUS=1 # Set to 8 for 32B and 70B models
MODEL=deepseek-ai/{model_name}
MODEL_ARGS="pretrained=$MODEL,dtype=bfloat16,max_model_length=32768,gpu_memory_utilisation=0.8,tensor_parallel_size=$NUM_GPUS"
OUTPUT_DIR=data/evals/$MODEL

lighteval vllm $MODEL_ARGS "custom|math_500|0|0" \
    --custom-tasks src/open_r1/evaluate.py \
    --use-chat-template \
    --output-dir $OUTPUT_DIR

Alternatively, you can launch Slurm jobs as follows:

python scripts/run_benchmarks.py --model-id={model_id}  --benchmarks math_500

GPQA Diamond

We are able to reproduce Deepseek's reported results on the GPQA Diamond benchmark within ~1-3 standard deviations:

Model	GPQA Diamond (🤗 LightEval)	GPQA Diamond (DeepSeek Reported)
DeepSeek-R1-Distill-Qwen-1.5B	33.3	33.8
DeepSeek-R1-Distill-Qwen-7B	48.4	49.1
DeepSeek-R1-Distill-Qwen-14B	55.6	59.1
DeepSeek-R1-Distill-Qwen-32B	58.6	62.1
DeepSeek-R1-Distill-Llama-8B	51.0	49.0
DeepSeek-R1-Distill-Llama-70B	65.2	65.2

To reproduce these results use the following command:

NUM_GPUS=1 # Set to 8 for 32B and 70B models
MODEL=deepseek-ai/{model_name}
MODEL_ARGS="pretrained=$MODEL,dtype=bfloat16,max_model_length=32768,gpu_memory_utilisation=0.8,tensor_parallel_size=$NUM_GPUS"
OUTPUT_DIR=data/evals/$MODEL

lighteval vllm $MODEL_ARGS "custom|gpqa:diamond|0|0" \
    --custom-tasks src/open_r1/evaluate.py \
    --use-chat-template \
    --output-dir $OUTPUT_DIR

python scripts/run_benchmarks.py --model-id={model_id}  --benchmarks gpqa

Data generation

Generate data from a smol distilled R1 model

The following example can be run in 1xH100. First install the following dependencies:

uv pip install "distilabel[vllm]>=1.5.2"

Now save the following snippet into a file named pipeline.py and run it with python pipeline.py. It will generate 4 outputs for each of the 10 examples (change the username for the repository to your org/user name):

from datasets import load_dataset
from distilabel.models import vLLM
from distilabel.pipeline import Pipeline
from distilabel.steps.tasks import TextGeneration


prompt_template = """\
You will be given a problem. Please reason step by step, and put your final answer within \boxed{}:
{{ instruction }}"""

dataset = load_dataset("AI-MO/NuminaMath-TIR", split="train").select(range(10))

model_id = "deepseek-ai/DeepSeek-R1-Distill-Qwen-7B"  # Exchange with another smol distilled r1

with Pipeline(
    name="distill-qwen-7b-r1",
    description="A pipeline to generate data from a distilled r1 model",
) as pipeline:

    llm = vLLM(
        model=model_id,
        tokenizer=model_id,
        extra_kwargs={
            "tensor_parallel_size": 1,
            "max_model_len": 8192,
        },
        generation_kwargs={
            "temperature": 0.6,
            "max_new_tokens": 8192,
        },
    )
    prompt_column = "problem"
    text_generation = TextGeneration(
        llm=llm, 
        template=prompt_template,
        num_generations=4,
        input_mappings={"instruction": prompt_column} if prompt_column is not None else {}
    )


if __name__ == "__main__":
    distiset = pipeline.run(dataset=dataset)
    distiset.push_to_hub(repo_id="username/numina-deepseek-r1-qwen-7b")

Take a look at the sample dataset at HuggingFaceH4/numina-deepseek-r1-qwen-7b.

Generate data from DeepSeek-R1

To run the bigger DeepSeek-R1, we used 2 nodes, each with 8×H100 GPUs using the slurm file present in this repo at slurm/generate.slurm. First, install the dependencies:

(for now we need to install the vllm dev wheel that fixes the R1 cuda graph capture)

pip install https://wheels.vllm.ai/221d388cc5a836fa189305785ed7e887cea8b510/vllm-1.0.0.dev-cp38-abi3-manylinux1_x86_64.whl --extra-index-url https://download.pytorch.org/whl/cu121

uv pip install "distilabel[vllm,ray,openai]>=1.5.2"

And then run the following command:

sbatch slurm/generate.slurm \
    --hf-dataset AI-MO/NuminaMath-TIR \
    --temperature 0.6 \
    --prompt-column problem \
    --model deepseek-ai/DeepSeek-R1 \
    --hf-output-dataset username/r1-dataset

Note

While the job is running, you can setup an SSH tunnel through the cluster login node to access the Ray dashboard from your computer running ssh -L 8265:ray_ip_head_node:8265 <login_node>, then browsing http://localhost:8265

Contributing

Contributions are welcome. Please refer to #23.