GanjinZero/wombat-7b-gpt4-delta

Model details

Organization developing the model
Alibaba DAMO Academy, Tsinghua University

Model date
Wombat-7B-GPT4 was released in 2023/04/13.

Model version
Wombat-7B-GPT4.

Training dataset
The training data of Wombat-7B-GPT4 is released in the GPT-4-LLM.

Model type
Wombat-7B-GPT4 is general-purpose instruction-following language model aligned with GPT4 (as proxy human preferences), fine-tuned from Alpaca models.
We use a novel method named RRHF (Rank Response to align Human Feedback) to fine-tune Alpaca.

How to use To recover Wombats from delta parameters:

python apply_delta.py \
    --base ./llama-7b \
    --target ./wombat-7b-gpt4 \
    --delta GanjinZero/wombat-7b-gpt4-delta

where apply_delta.py is from code.

To infer with Wombats: Please refer to code.

Citations details
Please cite our paper on Arxiv:

@misc{yuan2023rrhf,
      title={RRHF: Rank Responses to Align Language Models with Human Feedback without tears}, 
      author={Zheng Yuan and Hongyi Yuan and Chuanqi Tan and Wei Wang and Songfang Huang and Fei Huang},
      year={2023},
      eprint={2304.05302},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

License
Data are licensed under the CC BY NC 4.0 license.

Where to send questions or comments about the model
Questions, comments, and discussions about Wombats and RRHF can be sent via the GitHub repository of the project, by opening an issue.
or send emails to yuanzheng.yuanzhen@alibaba-inc.com, yuanhy20@mails.tsinghua.edu.cn or chuanqi.tcq@alibaba-inc.com.

Primary intended uses
The primary use of Wombat-7B and Wombat-7B-GPT4 is research on learning from human feedback and is a prototype of RRHF methods.

Primary intended users
The primary intended users of Wombat-7B and Wombat-7B-GPT4 are researchers in natural language processing, machine learning and artificial intelligence.

Out-of-scope use cases
Wombat-7B and Wombat-7B-GPT4 are not finetuned with proxy human feedback of OpenAI chatGPT and GPT4 and are not intended for use in production systems. Any usage must not compete with the OpenAI API.