Skip to content

Latest commit

 

History

History
131 lines (99 loc) · 4.04 KB

README.md

File metadata and controls

131 lines (99 loc) · 4.04 KB
Raw

Neuro-Symbolic Integration Brings Causal and Reliable Reasoning Proofs

This repo contains the code for the paper Neuro-Symbolic Integration Brings Causal and Reliable Reasoning Proofs

Table of Contents

Overview

CaRing (Causal and Reliable Reasoning) is a neuro-symbolic integration method for compelx reasoning problems. The main advantage of CaRing is that it produces strictly causal and reliable reasoning proofs along with the answer, allowing humans to understand the reasoning process of the model and verify the correctness and safety of the answer.

Environment Setup

1. Set Up Conda Environment (Required for prompting LLMs)

conda create -n caring python=3.9
conda activate caring

git clone https://github.com/DAMO-NLP-SG/CaRing.git
cd CaRing

# We use cuda=11.8, you may change it to your cuda version
pip install torch==2.0.1 torchvision==0.15.2 torchaudio==2.0.2 --index-url https://download.pytorch.org/whl/cu118

pip install -r requirements.txt

Here are some problems that you may encounter when installing the environment: [Issue-1].

2. Install SWI-Prolog (Required for evaluation)

Note: Below installation steps follow the guide from the SWI-Prolog official page.

  • For Ubuntu Machines:

    sudo apt-get update
sudo apt-get install software-properties-common

sudo apt-add-repository ppa:swi-prolog/stable
sudo apt-get update
sudo apt-get install swi-prolog

pip install git+https://github.com/yuce/pyswip@master#egg=pyswip

  • Other Linux Machines:

    See the SWI-Prolog official page.

Usage

Prompt LLMs

cd CaRing
mkdir output

export MODE=prolog  # Our Method
# export MODE=cot  # CoT

sh scripts/run_proofwriter.sh ${MODE}
# sh scripts/run_gsm8k.sh ${MODE}
# sh scripts/run_prontoqa.sh ${MODE}

You may customize the configurations in the corresponding .sh scripts. Currently we support Code-LLaMA series as the base LLM.

Evaluation:

  • ProofWriter 
    export MODE=prolog  # Our Method
# export MODE=cot  # CoT
python evaluate.py \
    --gpt_response_path ../output/proofwriter/${MODE}.CodeLlama-34b-hf.num_samples--1.demo-1_2.json \
    --dataset_name proofwriter \
    --evaluate_proof \
    --use_multiprocessing
  • GSM8K 
    export MODE=prolog  # Our Method
# export MODE=cot  # CoT
python evaluate.py \
    --gpt_response_path ../output/gsm8k/${MODE}.CodeLlama-34b-hf.num_samples--1.demo-1_2_3_4_5.json \
    --dataset_name gsm8k \
    --evaluate_proof \
    --use_multiprocessing
  • PrOntoQA 
    export MODE=prolog  # Our Method
# export MODE=cot  # CoT
python evaluate.py \
    --gpt_response_path ../output/prontoqa/${MODE}.CodeLlama-34b-hf.num_samples--1.demo-1_2.json \
    --dataset_name prontoqa \
    --use_multiprocessing

Citation

If you find this project useful, please consider citing our work:

@article{yang2023neuro,
  title={Neuro-Symbolic Integration Brings Causal and Reliable Reasoning Proofs},
  author="Yang, Sen  and
      Li, Xin and
      Cui, Leyang and
      Bing, Lidong and
      Lam, Wai",
  journal={arXiv preprint},
  year={2023}
}

Acknowledgments

This project makes use of the following open-source projects: