Unsupervised Calibration through Prior Adaptation

This is the code to reproduce the results of the paper Unsupervised Calibration through Prior Adaptation.

Installation

Install a conda environment to work with python 3.11.4:

conda create -n ucpa python=3.11.4
conda activate ucpa
pip install -r requirements.txt

Download the model weights:

python prepare_models.py --root_directory=. --model="gpt2-xl"

Usage

To run the code, use the following command:

bash run.sh

Code support GPU and CPU usage. It will try to allocate part of the model in the GPU, up to the specified size in the config file.

Citation

If you use this code, please cite the following paper:

@misc{estienne2023unsupervised,
      title={Unsupervised Calibration through Prior Adaptation for Text Classification using Large Language Models}, 
      author={Lautaro Estienne and Luciana Ferrer and Matías Vera and Pablo Piantanida},
      year={2023},
      eprint={2307.06713},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}