Open R1

This is a fully open-source replication project of the DeepSeek-R1 model, aimed at assisting developers in recreating and building models based on R1.

EditorRecommendationProductivityDeep LearningNatural Language Processing

Visit

huggingface/open-r1 is an open-source initiative dedicated to replicating the DeepSeek-R1 model. This project provides a range of scripts and tools for training, evaluating, and generating synthetic data, supporting various training methods and hardware configurations. Its primary advantage is its complete openness, allowing developers to freely use and improve it, making it a valuable resource for those looking to conduct research and development in deep learning and natural language processing. Currently, there is no specific pricing, making it suitable for both academic and commercial use.

Provides a complete training and evaluation process for the R1 model
including SFT and GRPO methods.
Supports multiple hardware configurations
such as DDP and DeepSpeed (ZeRO-2 and ZeRO-3).
Generates synthetic data using Distilabel
enriching training datasets.
Evaluates models with lighteval
supporting various tasks and model sizes.
Offers Makefile commands to simplify operations and enable users to quickly get started.

\This project is designed for developers
researchers
and enterprise users who wish to engage in research and development in the field of natural language processing. It offers a comprehensive framework that aids users in replicating and enhancing the DeepSeek-R1 model while supporting various hardware configurations and training methods
catering to projects of different scales and requirements.\

Fine-tune the model using the SFT method to tailor it for specific natural language processing tasks.
Enhance model performance on inference tasks through the GRPO method.
Utilize Distilabel to generate synthetic data
improving the model's generalization capability.

1. Create a Python virtual environment and install necessary dependencies such as vLLM and PyTorch.
2. Download the project code and configure the accelerator settings.
3. Train the model using either the SFT or GRPO scripts
adjusting parameters as needed.
4. Use the lighteval tool to evaluate model performance

Visit

Open R1 Visit Over Time

Monthly Visits

490881889

Bounce Rate

37.92%

Page per Visit

5.6

Visit Duration

00:06:18

Open R1 Visit Trend

Open R1 Visit Geography

Open R1 Traffic Sources

Open R1 Alternatives

Janus-Pro-1B — Janus-Pro-1B is an autoregressive framework for unified multi-modal understanding and generation.

Image•Multi-modal•Image Generation

DeepSeek-V3 — A Mixture-of-Experts language model with 671 billion parameters.

ChineseSelection•Natural Language Processing•Deep Learning

11634

RWKV — The new generation of large-scale model architecture, surpassing transformer.

Productivity•Open Source•Deep Learning

150

SpeechGPT 2.0-preview — The first human-level real-time interactive system focused on contextual intelligence, supporting multi-emotional and multi-style voice interactions.

chatting•Voice Interaction•Artificial Intelligence

YuE-s1-7B-anneal-en-cot — YuE is an open-source music generation model that transforms lyrics into complete songs.

Music•Music Generation•Deep Learning

Tarsier — Tarsier is a large video language model developed by ByteDance that generates high-quality video descriptions.

Video•Video Description•Video Understanding

432

leapfusion-hunyuan-image2video — A novel image-to-video sampling technology based on the Hunyuan model, enabling high-quality video generation.

Video•Artificial Intelligence•Video Generation

102

VideoLLaMA3 — VideoLLaMA3 is a cutting-edge multimodal foundational model focused on image and video understanding.

Video•Multimodal•Video Understanding

DeepSeek-R1-Distill-Qwen-1.5B — DeepSeek-R1-Distill-Qwen-1.5B is an efficient inference open-source language model suitable for various natural language processing tasks.

Programming•Natural Language Processing•Reinforcement Learning

1236

Flex.1-alpha — A pre-trained model for text-to-image generation with 8 billion parameters and licensed under Apache 2.0.

Image•Text-to-image generation•Deep learning

276

InternLM3-8B-Instruct — InternLM3-8B-Instruct is an open-source instruction model with 8 billion parameters designed for general-purpose use and advanced reasoning.

Programming•Large Language Model•Open Source

MiniMax-01 — A powerful language model with a total of 456 billion parameters, capable of processing context lengths of up to 4 million tokens.

Programming•Artificial Intelligence•Language Model

132

Llama-3.1-70B-Instruct-AWQ-INT4 — Text generation model with 70 billion parameters

Productivity•Text Generation•Natural Language Processing

VidTok — A family of open-source video segmenters from Microsoft.

Video•Video Segmentation•Video Compression

DRT-o1 — A deep inference translation model that enhances neural machine translation through extended reasoning chains.

Programming•Neural Machine Translation•Extended Reasoning Chains

120

Q-RWKV-6 32B Instruct Preview — The most powerful variant of the RWKV model, breaking multiple English benchmark tests.

Programming•Machine Learning•Natural Language Processing

mwp_ReFT — A deep reinforcement learning-based model fine-tuning framework

Programming•Natural Language Processing•Deep Learning

138

Florence-VL — Enhancement tool for visual language models, combining generative visual encoders and deep breadth fusion technology.

Programming•Visual Language Models•Multimodal Learning

108

PaliGemma 2 — PaliGemma 2 is a powerful visual language model that is easy to fine-tune.

Productivity•Visual Language Model•Machine Learning

OLMo-2-1124-7B-SFT — High-performance English text generation model

Productivity•Text Generation•Natural Language Processing

Sana — High-efficiency high-resolution image synthesis framework

Image•Image Synthesis•Text to Image

138

OLMo 2 7B — A large language model with 7 billion parameters, enhancing natural language processing capabilities.

Programming•Large Language Model•Natural Language Processing

Ai2 OpenScholar — Enhanced retrieval-based language model for scientific literature.

Productivity•Artificial Intelligence•Scientific Literature

Llama-3.1-Tulu-3-8B-RM — An advanced instruction-following model providing comprehensive guidance on post-training techniques.

Programming•Natural Language Processing•Post-Training Techniques

Llama-3.1-Tulu-3-70B-SFT — A leading family of instruction-following models, offering open-source data, code, and guidelines.

Programming•Natural Language Processing•Text Generation

Llama-3.1-Tulu-3-8B-DPO — An advanced text generation model that supports diverse tasks.

Programming•Text Generation•Natural Language Processing

playwright-ai — An AI tool for Playwright testing using the Anthropic API.

Programming•Playwright•Automation Testing

210

Llama-3.1-Tulu-3-8B — An advanced instruction-following model that provides open-source data and code.