We’ve been working with the community and collecting feedback to improve ChatLLaMA. With ChatLLaMA, you can create your hyper-personalized ChatGPT-like assistant using your own data and the least amount of compute possible.

Today we are happy to announce a new release of the project, with many new features and fixed issues 🚀

Main updates

• Now you can easily parallelize your training with HuggingFace Accelerate and DeepSpeed.
• You can efficiently fine-tune your Hugging Face's LLM with PEFT/LoRA.
• Use your dataset without worrying to break your training with automatic dataset checks.
• Your training stats are logged and easily accessible for maximum transparency of what is happening.
• Apply ChatLLaMA also to your favorite HuggingFace models for maximum flexibility.
• I want to align the assistant with my personal/company values, culture, brand and manifesto.
• Training is more robust with checkpoints automatically managed for you.
• Schedulers and Hyper-parameters are already initialized for an optimal starting point.

If you'd like to remain updated on our progress, join our Discord community.

‍