MuSR: Testing the Limits of Chain-of-thought with Multistep Soft Reasoning

Creating murder mysteries that require multi-step reasoning with commonsense using ChatGPT!

By: Zayne Sprague, Xi Ye, Kaj Bostrom, Swarat Chaudhuri, and Greg Durrett.

zayne@utexas.edu {xi, kaj, swarat, gdurrett}@cs.utexas.edu

View the dataset on our project website!

Check out the paper. Appearing at ICLR 2024 as a spotlight presentation!

MuSR Eval

The datasets are in datasets/{domain_name}.json

Install

1. virtualenv venv we have tested with python 3.8
2. source venv/bin/activate
3. pip install -r requirements.txt

Evaluate

To run the evaluation script on the MuSR datasets:

cd eval
OPENAI_API_KEY=key python eval.py

You can edit the functionality of the evaluation in eval.py as well (including different prompting strategies, models, and more).

[Optional] Install Redis for caching

We cache all LLM calls (openai and huggingface) with keys based on the prompt and model parameters to speed up evaluations.

To do this, we used Redis

Easiest way to install it is (for linux)

1. apt-get install redis
2. redis-server

Alternatively you can run our code without redis or disable the cache entirely by commenting out the lines cache.enable().

New models

Right now we support all the OpenAI endpoints and models published on Huggingface.

Custom models made in PyTorch or Tensorflow will need to have an implementation that follows from the Model class in src/model/model.py similar to src/model/hf.py (for Huggingface).

New prompts and MuSR domain datasets

These are easily added to the eval/eval.py file.

Overview of MuSR

This repository holds the code for the paper MuSR: Testing the Limits of Chain-of-thought with Multistep Soft Reasoning

All MuSR datasets can be found in {project_root}/datasets. Follow the installation guide to get the datasets mentioned from the paper downloaded.

Major components for making the MuSR dataset can be found in {project_root}/src.

Important classes and structures will have some example uses in their files (Madlib and LogicTree for example)

The scripts used to create a dataset can be found in {project_root}/musr_dataset_scripts

Evaluation scripts are in {project_root}/eval

Generating a dataset

Every dataset creation script is in {project_root}/musr_dataset_scripts. In those files are detailed instructions on how to create MuSR datasets as well as parameters for creating your own unique datasets! Individual components that are used to create each dataset should be heavily documented as well.

To run a script:

cd musr_dataset_scripts
OPENAI_API_KEY=key python {dataset_script}.py

NOTE: We tested most of this with GPT-4. It's possible that quality may significantly degrade if you use a different model due to the prompts being heavily tailored to GPT-4 as well as expecting the LLM to produce "good" outputs (some datasets require parsing of an output which requires strict formatting)

This will produce a dataset file in {project_root}/datasets after it completes.

Creating your own dataset

You can implement your own DatasetBuilder following the examples for the other domains.

For example, the important files used in creating murder mysteries are:

{project_root}/src/dataset_builder.py: The main file used for creating all datasets (shared functionality including the recursive reasoning tree expansion algorithm).

{project_root}/src/dataset_types/murder_mystery_dataset.py: Specific domain logic (and some prompts) for creating the murder mysteries.

{project_root}/musr_dataset_scripts/create_murder_mysteries.py: The main file that glues everything together (and includes some more prompts).