Agent-Tunix Documentation

Agent-Tunix is a GRPO (Group Relative Policy Optimization) training framework for Gemma3 language models using Google Tunix, with parameter-efficient fine-tuning via LoRA.

Getting Started

• Installation
  • System Requirements
  • CUDA and Driver Setup
  • Installation Methods
  • Verify Installation
  • Environment Variables
  • Troubleshooting
  • Next Steps
• Quick Start
  • Basic Training
  • View Configuration
  • Quick Testing
  • Custom Configuration
  • Using Experiment Presets
  • Model Evaluation
  • Utilities
  • Makefile Shortcuts
  • Hyperparameter Sweeps
  • Output Structure
  • Next Steps
  • Getting Help
• Configuration Guide
  • Configuration Overview
  • Configuration Structure
  • Main Configuration
  • Configuration Groups
  • Command-Line Overrides
  • Experiment Presets
  • Configuration Inspection
  • Advanced: Creating Custom Configurations
  • Configuration Best Practices
  • Memory Requirements by Model
  • Next Steps

User Guide

• Training Guide
  • Overview
  • Basic Training
  • Training Process
  • Training Configuration
  • Monitoring Training
  • Memory Optimization
  • Distributed Training
  • Resuming Training
  • Custom Reward Functions
  • Troubleshooting
  • Tips and Best Practices
  • Example Training Runs
  • Next Steps
• Evaluation Guide
  • Model Evaluation
  • Evaluation Metrics
  • Configuration
  • Inference Strategies
  • Multiple Passes
  • Checkpoint Selection
  • Advanced Configuration
  • Batch Evaluation
  • Interpreting Results
  • Troubleshooting Evaluation
  • Example Evaluation Workflow
  • Programmatic Evaluation
  • Next Steps
• Hyperparameter Tuning
  • Overview
  • Key Hyperparameters
  • Tuning Strategies
  • Monitoring During Tuning
  • Common Tuning Issues
  • Efficient Search
  • Memory-Aware Tuning
  • Example Tuning Workflow
  • Best Practices
  • Next Steps
• Experiments Guide
  • What are Experiments?
  • Using Built-in Experiments
  • Creating Custom Experiments
  • Experiment Structure
  • Example Experiments
  • Experiment Workflow
  • Combining Experiments with Sweeps
  • Advanced: Experiment Hierarchies
  • Naming Conventions
  • Best Practices
  • Managing Experiments
  • Next Steps

Configuration

• Configuration Overview
  • Directory Structure
  • Main Configuration (config.yaml)
  • Configuration Groups
  • Using Configurations
  • Configuration Composition
  • Interpolations
  • Creating Custom Configurations
  • Configuration Validation
  • Best Practices
  • Configuration File Format
  • Environment Variables
  • Common Patterns
  • Next Steps
• Model Configuration Reference
  • Available Models
  • Model Configuration Parameters
  • Complete Model Configuration Example
  • Memory Requirements by Configuration
  • Tuning LoRA Rank
  • Creating Custom Models
  • Integration with Training
  • Next Steps
• Optimizer Configuration Reference
  • Available Optimizers
  • Optimizer Configuration Parameters
  • Learning Rate Scheduling
  • Optimizer Tuning Workflow
  • Common Configuration Examples
  • Diagnosing Optimizer Issues
  • Complete Optimizer Configuration Example
  • Interaction with Other Settings
  • Next Steps
• Training Configuration Reference
  • Training Configuration File
  • Training Configuration Parameters
  • Memory Optimization Parameters
  • Complete Training Configuration Example
  • Common Configuration Patterns
  • Resuming Training
  • Training Dynamics
  • Checkpoint Management
  • Monitoring Training Progress
  • Integration with Other Configs
  • Next Steps

API Reference

• Training API
  • Main Entry Point
  • Training Utilities
  • Configuration Classes
  • Training Flow
  • GRPO Algorithm
  • Example Usage
  • Tips and Best Practices
  • Next Steps
• Evaluation API
  • Main Entry Point
  • Evaluation Functions
  • Configuration Classes
  • Inference Configurations
  • Evaluation Metrics
  • Checkpoint Selection
  • Advanced Configuration
  • Multiple Passes
  • Batch Evaluation
  • Interpreting Results
  • Troubleshooting
  • Programmatic Usage
  • Example Evaluation Workflow
  • Next Steps
• Models API
  • Available Models
  • Model Configuration
  • LoRA (Low-Rank Adaptation)
  • Custom Model Configuration
  • Distributed Training
  • Model Loading
  • Memory Requirements by Model
  • Advanced Topics
  • Next Steps
• Data API
  • Dataset Support
  • Data Loading
  • Custom Datasets
  • Data Format Requirements
  • Tokenization
  • Preprocessing Pipeline
  • Data Splits
  • Batch Processing
  • Sequence Lengths
  • Advanced: Custom Data Loading
  • Data Validation
  • Troubleshooting
  • Next Steps
• Rewards API
  • Reward System Overview
  • Built-in Reward Functions
  • Combined Reward
  • Custom Reward Functions
  • Reward Design Patterns
  • Reward Shaping
  • Reward Debugging
  • Common Reward Issues
  • Testing Rewards
  • Advanced Topics
  • Next Steps

Advanced Topics

• Distributed Training
  • Overview
  • Configuration
  • Single GPU Setup
  • Data Parallel (Multiple GPUs)
  • Tensor Parallel (Model Sharding)
  • Hybrid Parallelism
  • Multi-Node Setup
  • Performance Tuning
  • Debugging Distributed Training
  • Common Issues
  • Example Distributed Setups
  • Monitoring
  • Best Practices
  • Next Steps
• Custom Reward Functions
  • Basic Reward Function Structure
  • Implementing Reward Functions
  • Registering Custom Rewards
  • Iterating on Rewards
  • Reward Normalization
  • Reward Shaping for Better Training
  • Common Issues and Solutions
  • Debugging Reward Functions
  • Best Practices
  • Next Steps
• Troubleshooting
  • Installation Issues
  • Training Issues
  • Evaluation Issues
  • Configuration Issues
  • Distributed Training Issues
  • Debugging Tips
  • Getting Help
  • Common Patterns
  • Next Steps

References

• Frequently Asked Questions
• Glossary

Indices and tables

• Index

• Module Index

• Search Page

Quick Links

• Quick Start Guide

• Configuration Guide

• API Reference

• GitHub Repository