Repo for the paper PaL: Program-Aided Language Models.
In PaL, Large Language Model solves reasoning problems that involve complex arithmetic and procedural tasks by generating reasoning chains of text and code. This offloads the execution of the code to a program runtime, in our case, a Python interpreter. In our paper, we implement PaL using a few-shot prompting approach.
This repo provides an interactive implementation of PAL.
[Mar 2023] We have added supports for ChatGPT APIs (e.g., gpt-3.5-turbo). We expect a smooth transition for PAL over the codex API shutdown. Checkout a beta script scripts/gsm_chatgpt.py
for Math reasoning.
[Jan 2023] We release GSM-hard, a harder version of GSM8k we created. Also avaliable on Huggingface 🤗
import datasets
gsm_hard = datasets.load_dataset("reasoning-machines/gsm-hard")
Clone this repo and install with pip
.
git clone https://github.com/luyug/pal
pip install -e ./pal
Before running the scripts, set the OpenAI key,
export OPENAI_API_KEY='sk-...'
The core components of the pal
package are the Interface classes. Specifically, ProgramInterface
connects the LLM backend, a Python backend and user prompts.
import pal
from pal.prompt import math_prompts
interface = pal.interface.ProgramInterface(
model='code-davinci-002',
stop='\n\n\n', # stop generation str for Codex API
get_answer_expr='solution()' # python expression evaluated after generated code to obtain answer
)
question = 'xxxxx'
prompt = math_prompts.MATH_PROMPT.format(question=question)
answer = interface.run(prompt)
Here, the interface
's run
method will run generation with the OpenAI API, run the generated snippet and then evaluate get_answer_expr
(here solution()
) to obtain the final answer.
User should set get_answer_expr
based on the prompt.
We provide simple inference loops in the scripts/
folder.
mkdir eval_results
python scripts/{colored_objects|gsm|date_understanding|penguin}_eval.py
We have a beta release of a ChatGPT dedicated script for math reasoning.
python scripts/gsm_chatgpt.py
For running bulk inference, we used the generic prompting library prompt-lib and recommend it for running CoT inferenence on all tasks used in our work.
For the complete details of the results, see the paper .
@article{gao2022pal,
title={PAL: Program-aided Language Models},
author={Gao, Luyu and Madaan, Aman and Zhou, Shuyan and Alon, Uri and Liu, Pengfei and Yang, Yiming and Callan, Jamie and Neubig, Graham},
journal={arXiv preprint arXiv:2211.10435},
year={2022}
}