Model Overview

GanjinZero/wombat-7b-delta is a 7 billion parameter instruction-following language model, developed by Alibaba DAMO Academy and Tsinghua University. Released on April 13, 2023, it is fine-tuned from Alpaca models using a novel method called RRHF (Rank Response to align Human Feedback). This technique aligns the model with human preferences, using ChatGPT as a proxy for human feedback.

Key Capabilities

• Instruction Following: Designed to follow instructions effectively.
• Human Feedback Alignment: Utilizes the RRHF method for alignment, a key research focus.
• Research Prototype: Serves as a prototype for the RRHF methodology.

Intended Use Cases

• Research on Human Feedback: Primarily intended for researchers studying methods of learning from human feedback in language models.
• RRHF Method Exploration: Ideal for exploring and understanding the RRHF fine-tuning approach.

Limitations

• Not for Production: This model is not fine-tuned for production systems and is not intended for commercial use.
• No Competition with OpenAI API: Usage must not compete with the OpenAI API.