This repository includes the source code for our paper Reverse Thinking Makes LLMs Stronger Reasoners.

Installation

This repository has been tested with Python 3.10.14. We recommend using Conda for environment management. You can set up the Conda environment using the following commands:

conda create --name revthink python=3.10.14 -y
conda activate revthink
pip install -r requirements.txt

Run Experiments

Args usages

• --task: which dataset to use, can be: SQA, CSQA, ARC, MATH, GSM8K, TabMWP, ANLI, Date.
• --n: the proportion of training data to use, can be 1-100. Default: 100.
• --model: the student model to train, can be: mistral-7b, gemma-7b.

Extract data

All data files can be download from https://drive.google.com/file/d/1zoi4KNb9vzl8vd8Xh0dHfJ7DwSQDiiK7/view?usp=share_link, which includes a data.zip with two directories: /training_data and /test_data. Extract them using the following command:

unzip data.zip

We use the dataset Date as an example below.

Run

Step 1 (Optional): Use the following script to augment the original dataset by generating forward reasoning, backward questions, and backward reasoning using the teacher model, Gemini-1.5-Pro:

python augment_data.py --task SQA

If you wish to generate the data by yourself, you will need to set up Vertex AI following this tutorial, and fill in the project_id and location in utils.py.

Note: This step is optional as we provide the augmented datasets in the /training_data directory.

Step 2: Train the student model using the following script:

CUDA_VISIBLE_DEVICES=0 python train.py --task SQA --n 100 --model mistral-7b

Step 3: Evaluate the student model with the following script:

CUDA_VISIBLE_DEVICES=0 python evaluate.py --task SQA --n 100 --model mistral-7b

For out-of-domain (OOD) evaluation, you can use the following tasks as the --task argument: BoolQ, OBQA, ESNLI, or GSM8K-Rev.