TextRL

About

A reinforcement learning library for fine-tuning text generation models using RL techniques like PPO, enabling alignment with human preferences.

• Implements Proximal Policy Optimization (PPO) for text generation
• Supports multiple transformer models (GPT-2, GPT-Neo, BlenderBot, etc.)
• Built-in reward model training from human feedback
• Easy integration with Hugging Face Transformers
• Customizable reward functions and training loops