Understand the project
Practical overview
When it fits
Use EasyR1 when you need a code-based workflow for reinforcement-learning training of supported language or multimodal models, and can manage the required Python, GPU and related software environment.
Who it is for
- AI researchers training or adapting language and vision-language models with reinforcement-learning methods.
- Developers preparing text, image-text or mixed-modality datasets for model training.
- Teams configuring GPU environments or Ray-based multi-node training workflows.
Documented uses
- Train listed language or vision-language models using supported reinforcement-learning algorithms.
- Prepare custom text, image-text, multi-image-text or mixed text-image datasets in the required format.
- Use LoRA or full fine-tuning examples and resume from checkpoints.
- Merge a training checkpoint into Hugging Face format and track runs with supported logging tools.
Before you choose
Requirements & limits
- The framework is based on veRL and uses Python; documented software prerequisites include Python 3.9+, Transformers 4.54.0+, Flash-Attention 2.4.3+ and vLLM 0.8.3+.
- Supported model families and algorithms are explicitly listed, but model integrations beyond those lists should not be assumed.
- Dataset support includes text and several image-text forms, provided the data follows the required format.
- The README’s estimated GPU-memory requirements vary considerably with model size, precision and full versus LoRA fine-tuning.
- Docker is supported through a supplied Dockerfile and recommended pre-built image; Apptainer is described as an alternative where Docker is unavailable.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Source-based practical guide
Install EasyR1 and launch the Geometry3K GRPO example
Set up the repository and run its documented Qwen2.5-VL Geometry3K GRPO training script.
Before you start
- Python 3.9 or later
- Transformers 4.54.0 or later, Flash-Attention 2.4.3 or later, and vLLM 0.8.3 or later
- An environment with suitable GPU resources; the README provides estimated requirements by model size and method
Steps
Clone the EasyR1 repository and change into its directory.
Install the project with the command below.
pip install -e .Launch the documented example with the command below.
bash examples/qwen2_5_vl_7b_geo3k_grpo.sh
Expected outcome
The documented Qwen2.5-VL GRPO training workflow for the Geometry3K dataset is launched.
Guide limitations
- The README’s hardware figures are estimates; resource needs depend on model size and training method.
- This recipe follows the documented commands but does not establish that a given local environment will complete training.
Evidence and transparency
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 7 October 2026
- Original website title
- hiyouga/EasyR1 · Official homepage title · retained in its source language