Home

Awesome

PRD: Peer Rank and Discussion Improve Large Language Model based Evaluations

| Paper | Demo |

Peer Rank (PR) process:

Peer Rank (PR) process

Peer Discussion (PD) process:

Peer Discussion (PD) process

Install Dependencies

Please follow the following commands to install dependencies.

# create an environment
conda create -n prd python=3.8
conda activate prd

# install from the requirement file by pip
pip install -r requirement.txt

Datasets

We publish the dataset Vicuna80 in the data folder. For information about datasets, please refer to the README file.

Generated Results

For information about generated results, please refer to the README file.

Run

Please follow the bash commands to run corresponding parts.

Peer Rank

Please enter the peer_rank folder by the following command.

cd peer_rank/

Reviews Generation

Please run the gen_{reviewer}.sh scripts to generate reviews for answers from one pair of model. For example,

./gen_claude.sh ../data/vicuna80/generations/answer_[Model 1].jsonl ../data/vicuna80/generations/answer_[Model 2].jsonl

To generate reviews for answers from all pairs of models, please run the gen_{reviewer}_all.sh. For example,

./gen_claude_all.sh

Peer Ranking

To run peer ranking, please open the peer_ranking.ipynb file by any Jupyter Notebook.

Peer Discussion

Please enter the peer_discussion folder by the following command.

cd peer_discussion/

Before running any python script, please make sure the file config.yml contains correct configurations you need.

Reviews Generation

python review_lfqa.py

There is no codes of generating reviews for Vicuna80 since they are provided in the Peer Rank related codes.

Discussion Generation

# discuss on LFQA
python gather_all_lfqa.py
python discuss_lfqa.py

# discuss on Vicuna80
python gather_all_vicuna80.py
python discuss_vicuna80.py

Citation

Please cite the following if find our work helpful.

@misc{li2023prd,
      title={PRD: Peer Rank and Discussion Improve Large Language Model based Evaluations},
      author={Ruosen Li and Teerth Patel and Xinya Du},
      year={2023},
      eprint={2307.02762},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

Contact

Following 2 options are available for any clarification, comments or suggestions