LLM-Dojo

LLM-Dojo

Welcome to LLM-Dojo, an open-source training ground for large models. This project emphasizes simplicity and readability in its code base while providing robust functionalities for training various mainstream models such as Qwen, Llama, and GLM. The framework also includes Reinforcement Learning from Human Feedback (RLHF) methods like DPO, CPO, KTO, and PPO.

Key Features:

• Versatile Framework: Supports both language (LLM) and vision models (VLM).
• Multiple Training Methods: Provides support for state-of-the-art RLHF methods, knowledge distillation, and rejected sampling.
• Flexibility: Automatically adapts training templates based on the provided data, streamlining the training process.
• High Compatibility: Supports multi-GPU setups through Deepspeed and allows for Lora, QLora, and Dora methods for fine-tuning.

Benefits:

• User-Friendly: Designed with an emphasis on code clarity, making it easy for developers to learn and customize.
• Community-Driven: Continuous updates and improvements based on user feedback contribute to its evolving capabilities.

Highlights:

• A dedicated section for RLHF methodologies that are continually updated.
• Clear documentation and examples for users to quickly get started.
• Active community involvement through issues and pull requests on GitHub.

Start your journey in large model training with LLM-Dojo, where learning and experimentation are encouraged!