PINTO

This is a Pytorch implementation for our ICLR 2023 paper: Faithful Language Reasoning Using Prompt-Generated Rationales [arxiv].

Prepare data

We have provided the datasets augmented with prompt-generated rationales. Just untar the file data.tar.gz. Since our data is based on existing benchmarks, please cite their works if you use the data.

Alternaltively, you can use the code for generating rationales rationalization_prompting.py to prepare the rationales for your own datasets.

Training

Run the script run.sh. Change the dataset argument to specify the dataset for experiment. After training, the evaluation result is saved to ./checkpoint.

Citation

@inproceedings{
wang2023pinto,
title={{PINTO}: Faithful Language Reasoning Using Prompted-Generated Rationales},
author={PeiFeng Wang and Aaron Chan and Filip Ilievski and Muhao Chen and Xiang Ren},
booktitle={International Conference on Learning Representations},
year={2023},
url={https://openreview.net/forum?id=WBXbRs63oVu}
}