Home

Awesome

CARP

This repo provides the CARP dataset and the source code of our paper: Evaluating and Improving Tool-Augmented Computation-Intensive Math Reasoning.

Introduction

We construct CARP, a new Chinese dataset consisting of 4,886 computation-intensive algebra problems with formulated annotations on intermediate steps. Furthermore, we propose a new approach that can deliberate the reasoning steps with tool interfaces, namely DELI. deli_model

The CARP Dataset

The CARP dataset consists of 4,886 middle school computation-intensive algebra problems, and each problem is associated with a natural language solution and an annotated EFG. Our annotated EFG explicitly depicts the step-by-step reasoning process of a math problem in a readable and concise format.

carp_desc carp_case

The CARP dataset and other datasets evaluated in the paper are stored in data directory.

Quick Start

Setup

  1. Install dependencies.
conda create --name carp python=3.9
conda activate carp
pip install -r requirements.txt
  1. Create a file ./openai_api_key and put your OpenAI API keys in it.

Run

CARP

# Random CoT
bash scripts/carp/random_cot.sh
# Retrieval CoT
bash scripts/carp/retrieval_cot.sh
# DELI
bash scripts/carp/deli.sh

MATH

DATA_NAME="algebra" # prealgebra / number_theory / counting_and_probability

# Random CoT
bash scripts/MATH/random_cot-general.sh $DATA_NAME
# Retrieval CoT
bash scripts/MATH/retrieval_cot-genral.sh $DATA_NAME
# DELI
bash scripts/MATH/deli-general.sh $DATA_NAME

AGIEval

# Gaokao-MathCloze
# Random-CoT
bash scripts/agieval/gaokao-mathcloze/random_cot.sh
# DELI
bash scripts/agieval/gaokao-mathcloze/deli.sh

# SAT-Math
# Random-CoT
bash scripts/agieval/sat-math/random_cot.sh
# DELI
bash scripts/agieval/sat-math/deli.sh

License

The CARP datasets in data/carp and data/carp_en are under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.

CC BY-NC-SA 4.0

Citation

Please consider citing our work if you find it useful.

@article{zhang2023evaluating,
  title={Evaluating and Improving Tool-Augmented Computation-Intensive Math Reasoning},
  author={Zhang, Beichen and Zhou, Kun and Wei, Xilin and Zhao, Wayne Xin and Sha, Jing and Wang, Shijin and Wen, Ji-Rong},
  journal={arXiv preprint arXiv:2306.02408},
  year={2023}
}

Reference

We refer to the following repos: