Plausibility Vaccine

Injecting LLM Knowledge for Event Plausibility

Read the paper»

About The Project

Plausibility Vaccine is a library that investigates parameter-efficient finetuning and modular transfer of prompted physical property knowledge for modelling event plausibility.

Getting Started

Prerequisites

The project uses uv to manage and lock project dependencies for a consistent and reproducible environment. If you do not have uv installed on your system, visit this page for installation instructions.

Note: If you have pip you can just invoke:

pip install uv

Installation

# Clone the repo
git clone https://github.com/Jacob-Chmura/plausibility-vaccine.git

# Enter the repo directory
cd plausibility-vaccine

# Install core dependencies into an isolated environment
uv sync

# [Optional] Install extra dependencies to run analytics
uv sync --group analytics

Usage

Running Plausibility Vaccine

Full End-to-End Experiments

./run_plausibility_vaccine.sh

Baseline Experiments Only

./run_plausibility_vaccine.sh config/albert_reduce_factor_64/baseline.yaml

Pre-training Adapters Only

./run_plausibility_vaccine.sh config/albert_reduce_factor_64/pretraining.yaml

Running Analytics

All Analytics

Note: requires that you have previously ran plausibility_vaccine.sh and have generated results

./run_analytics.sh

Non-result Dependent Analytics

./run_analytics.sh --no-results

License

Distributed under the MIT License. See LICENSE.txt for more information.

Contact

Jacob Chmura - jacobpaul.chmura@gmail.com

Citation

@article{chmura-etal-2024-plausibility,
  title   = "Plausibility Vaccine: Injecting LLM Knowledge for Event Plausibility",
  author  = "Chmura, Jacob and Dauvet, Jonah, and Sabry, Sebastian"
  journal = "arXiv preprint arXiv:2503.12667",
  url     = "https://arxiv.org/abs/2503.12667",
  year    = "2024",
}

(back to top)