project-baize/baize-v2-13b

TEXT GENERATIONConcurrency Cost:1Model Size:13BQuant:FP8Ctx Length:4kPublished:May 23, 2023License:cc-by-nc-4.0Architecture:Transformer0.0K Open Weights Cold

Baize-v2-13b is a 13 billion parameter open-source chat model developed by UCSD and Sun Yat-Sen University, fine-tuned with LoRA using supervised fine-tuning (SFT) and self-distillation with feedback (SDF). This model is designed for conversational AI, providing detailed responses in a specific chat format. It is merged with LLaMA, making it ready for direct use in chat applications.

Loading preview...

Baize-v2-13b: An Open-Source Chat Model

Baize-v2-13b is a 13 billion parameter open-source chat model developed by UCSD and Sun Yat-Sen University. It is fine-tuned using LoRA with a combination of supervised fine-tuning (SFT) and self-distillation with feedback (SDF) techniques. The model is named after Baize, a mythical Chinese creature known for its extensive knowledge and ability to speak human languages, reflecting the developers' aspirations for its conversational capabilities.

Key Capabilities & Features

  • Conversational AI: Specifically designed and fine-tuned for chat-based interactions.
  • Detailed Responses: Engineered to provide comprehensive and detailed answers.
  • Specific Chat Format: Requires a precise input format ([|Human|] and [|AI|]) for optimal performance, ensuring structured dialogue.
  • Ethical Guidelines: Programmed to decline engagement with unethical, controversial, or sensitive topics.
  • LLaMA Integration: The checkpoint is merged with LLaMA, making it readily usable.

Good For

  • Developers building chat applications requiring a structured conversational model.
  • Research into self-distillation and LoRA fine-tuning methods for large language models.
  • Creating AI assistants that provide detailed, markdown-formatted responses while adhering to ethical content guidelines. More details and usage instructions can be found on the Project Baize GitHub and in their research paper.