Skip to content
/ REAR Public
forked from yhao-wang/REAR

Implementation of "REAR: A Relevance-Aware Retrieval-Augmented Framework for Open-Domain Question Answering"

Notifications You must be signed in to change notification settings

RUCAIBox/REAR

 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

12 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

REAR (EMNLP 2024)

REAR is a RElevance-Aware Retrieval-augmented framework for open-domain question answering. [paper]

The checkpoint is availible on huggingface🤗. [checkpoint]

🚀 Quick Start

  1. Install dependencies.

    pip install -r requirements.txt
  2. Run the following codes:

    from rear.src.inf import get_vllm_model
    from rear.src.routing import reliability
    get_vllm_model("RUCAIBox/rear-llama-7b-hf")
    example = {
        "question": "Who won the first Noble Prize in Physics",
        "ctxs": [
            "Wilhelm Conrad Röntgen won first Nobel Prize in Physics.",
            "Wilhelm Conrad Röntgen won it for discovery of X-rays",
            "Albert Einstein was awarded the 1921 Nobel Prize in Physics",
            "The Nobel Prize in Physics is a yearly award.",
            "First law of thermodynamics was stated by William"
            ]
        }
    final_answer = reliability(example)['rely_answer']
    print(final_answer)

🔍 Training and Inference Scripts

  1. Training.

    bash train.sh meta-llama/Llama-2-7b-hf [model output path] [training data] [deepspeed config file]
  2. Inference.

    First, to generate answers and the path-reliability scores:

    python rear/inference.py \
        --model_path RUCAIBox/rear-llama-7b-hf \
        --phase reliability \
        --source [test file] \
        --outfile [output file(1)]

    Second, to generate the knowledge-consistency scores:

    python rear/inference.py \
        --model_path RUCAIBox/rear-llama-7b-hf \
        --phase consistency \
        --source [output file(1)] \
        --outfile [output file(2)]

    The QA pairs, along with the documents used for inference, are available for download at the following link: [data].

    After running these scripts, if you have provided "reference" as the ground truth in the test source data, the EM (Exact Match) and F1 scores will be automatically calculated.

🌟 Acknowledgement

Please cite the following paper if you find our code or data helpful.

@article{wang2024rear,
    title={REAR: A Relevance-Aware Retrieval-Augmented Framework for Open-Domain Question Answering},
    author={Wang, Yuhao and Ren, Ruiyang and Li, Junyi and Zhao, Wayne Xin and Liu, Jing and Wen, Ji-Rong},
    journal={arXiv preprint arXiv:2402.17497},
    year={2024}
}

About

Implementation of "REAR: A Relevance-Aware Retrieval-Augmented Framework for Open-Domain Question Answering"

Resources

Stars

Watchers

Forks

Languages

  • Python 97.9%
  • Shell 2.1%