What problem does it solve?
This Skill automates the process of aligning Large Language Models (LLMs) with human preferences and instructions, making them more helpful, harmless, and honest.
Core Features & Use Cases
- Reinforcement Learning from Human Feedback (RLHF): Implements full RLHF pipelines including Supervised Fine-Tuning (SFT), Reward Model training, and Proximal Policy Optimization (PPO).
- Direct Preference Optimization (DPO): Aligns models directly using preference data without needing a separate reward model, simplifying the alignment process.
- Memory-Efficient RL: Utilizes Group Relative Policy Optimization (GRPO) for training with reduced memory footprints.
- Use Case: You have a base LLM that generates factually correct but sometimes unhelpful or biased responses. Use this Skill to fine-tune it using human-ranked responses, resulting in a model that better follows instructions and aligns with desired behavior.
Quick Start
Use the fine-tuning-with-trl skill to perform supervised fine-tuning on the 'Qwen/Qwen2.5-0.5B' model using the 'trl-lib/Capybara' dataset.