REGAL: Refactoring Programs to Discover Generalizable Abstractions

• Authors: Elias Stengel-Eskin, Archiki Prasad, Mohit Bansal
• Paper

Overview

This repo contains code for our paper: REGAL: Refactoring Programs to Discover Generalizable Abstractions REGAL is a method for learning libraries of interpretable helper functions for program synthesis in a diverse range of tasks. Please see the figure below (and our paper) for more details.

Dependencies

Dependencies can be installed by running

pip install -r requirements.txt 

Scripts

Scripts for preprocessing ReGAL can be found in scripts/<domain>/preprocess.sh. These require an OPENAI_API_KEY environment variable to be set.

Scripts for training ReGAL on LOGO and Date can be found in scripts/<domain>/train_regal.sh After training, the resulting logs and saved CodeBank files will be in logs.

TextCraft: coming soon!

Scripts for evaluation on CodeLlama-13B are in scripts/<domain>/test_regal.sh

Reference

Please cite our paper as

@article{stengeleskin2024regal,
    title = {ReGAL: Refactoring Programs to Discover Generalizable Abstractions},
    author = {Stengel-Eskin, Elias and Prasad, Archiki and Bansal, Mohit}, 
    year = {2024},
    journal={arXiv preprint 2401.16467} 
}