Fine-tune `Llama 3.1` 8B with `ORPO`

[Haleshot]

📋 Overview

Aiming to implement a unified fine-tuning approach for Llama 3 using ORPO (Odds Ratio Preference Optimization), showcasing it through interactive marimo notebooks. ORPO combines SFT + DPO into a single efficient process, making it more resource-efficient while maintaining or improving performance.

🎯 Objectives

• Implement ORPO fine-tuning for Llama 3.1 8B model using the TRL library

🔍 Technical Details

ORPO Implementation

• Combine instruction tuning and preference alignment in a single stage
• Use TRL library for implementation
• Target model: Llama 3 8B
• Based on approaches from:
  • MLAbonne's ORPO Guide

💡 Additional Contributions

• Plans to create additional demos using Marimo for other repository tasks/issues
• Complement existing Gradio demos with Marimo alternatives

📝 Todo List

• Set up initial marimo notebook structure
• Implement ORPO fine-tuning pipeline

🤝 Related Issues

• References Call for Contributions (Call for contributions #43)
• Complements the Gradio Demos initiative (Building Gradio Demos with Llama  #44)

📚 Resources

• Marimo.io documentation
• TRL library documentation

Looking forward to feedback and suggestions from the community!