Contributing to Trinity-RFT

Thank you for your interest in Trinity-RFT! Our framework is built on a decoupled architecture consisting of the Explorer, Trainer, and Buffer. We welcome all forms of contributions—from core feature enhancements and new algorithms to documentation and bug reports.

Where to Contribute

Trinity-RFT provides modular interfaces for different technical interests. Please refer to our Developer Guide for detailed implementation standards:

Focus Area	Interface/Code Directory	Potential Tasks
Agentic Workflows	Workflow	Implementing multi-turn dialogs, ReAct workflows, or domain-specific agent training capabilities (e.g., Coding, Math).
RL Algorithms	Algorithm	Integrating new RL algorithms (e.g., RLOO, GSPO) or optimizing loss functions and advantage estimations.
Data & Experience	Operator, Selector	Designing data cleaning, selection, reward modeling, or experience augmentation and replay strategies.
Use Cases	Examples	Sharing new usages and improving built-in demonstrated configurations.
General Utility	docs/, tests/	Improving documentation, adding translations, fixing bugs, or enhancing CLI/GUI tools.

How to Start: The "Plugin-First" Approach

To minimize friction and keep the core codebase stable, we recommend the Plugin-First workflow for new features:

1. Develop: Create your custom module in the trinity/plugins/ directory.
2. Auto-Load: Trinity-RFT automatically detects and registers modules in this directory at runtime without requiring changes to the framework's internal code.
3. Integrate: Once your feature is stable and verified, submit a Pull Request to "graduate" your code from plugins/ into the formal modules (e.g., trinity/algorithm/ or trinity/common/workflows/).

Submission Checklist

To ensure a smooth review process, please complete the following:

1. Registration: If moving code from a plugin to the core framework, register it in the corresponding __init__.py mapping.
2. Testing: Add or update unit tests in the tests/ directory. Verify your changes by running:

   python -m pytest tests/
   

3. Code Style: We use pre-commit to maintain code quality. Run the following before committing:

   pre-commit run --all-files
   

4. Description: Provide a clear PR title and a description that explains the motivation (why this change is needed) and the implementation (how it works).

---

Additional Guidelines

• Bug Reports & Feature Requests: Please use GitHub Issues. For bugs, include reproduction steps, environment info, and error logs.
• Major Changes: For significant architectural changes or large features, please open an issue first to discuss the design with the maintainers.
• Documentation: We highly value improvements to our tutorials, docstrings, and translations.

For a deep dive into the framework's architecture, please refer to the Full Doc.

Thank you for helping us build a better Reinforcement Fine-Tuning framework!