easy-r1 is an open-source project whose goal is to provide developers with a convenient way to train reasoning large language models similar to DeepSeek R1.
Purpose: This project focuses on the training of reasoning LLMs, helping users reproduce or experiment with a reasoning-capability training pipeline similar to DeepSeek R1 with a low barrier to entry. The project uses Qwen2.5-0.5B as the base model for training, and clearly states that it uses the Base model rather than the Instruct model.
Main capabilities: The project provides simplified training scripts and workflows, enabling users to start training by executing commands according to the instructions. During training, SwanLab is integrated for experiment tracking and recording. Users need to register a SwanLab account and obtain an API Key from the settings page before they can log in and use it. The project instructions provide a command-line login method, and also support logging in via the swanlab.login() function.
How to use: Before running the training command, the following prerequisites must be met:
- Install Python 3.10 or later;
- Use a computer equipped with an Nvidia GPU;
- Have CUDA and cuDNN installed. After that, register a SwanLab account, log in, copy the API Key from the settings page, paste it into the command line, and press Enter to complete login (the swanlab.login() function can also be used). Once ready, training for Qwen2.5-0.5B can be started according to the project instructions.
Applicable scenarios: Suitable for individual developers, researchers, and students who want to study or experiment with reasoning model training methods, especially scenarios where they want to try a DeepSeek R1-like reasoning training pipeline at low cost with a smaller model (Qwen2.5-0.5B).
Limitations: The project explicitly requires an Nvidia GPU and a CUDA, cuDNN environment, so corresponding hardware and software support are needed. The base model used for training is the Qwen2.5-0.5B Base model. Specific information such as the training dataset, training duration, VRAM requirements, and model performance evaluation is not currently provided in the available materials.
