Can you rely on your model evaluation? Improving model evaluation with synthetic test data

This repository contains code for the paper "Can you rely on your model evaluation? Improving model evaluation with synthetic test data"

For more details, please read our NeurIPS 2023 paper

Installation

1. Clone the repository
2. Create a new conda environment with Python 3.7. e.g:

    conda create --name 3s_env python=3.7

3. Install requirements in env

    pip install -r requirements.txt

4. Link the venv to the kernel:

    python -m ipykernel install --user --name=3s_env

Use-cases

We highlight different use-cases of 3S-Testing for both subgroup and shift testing in notebooks which can be found in the /use_cases folder.

Citing

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

@inproceedings
{3STesting,
title={Can you rely on your model evaluation? Improving model evaluation with synthetic test data},
author={van Breugel, Boris and Seedat, Nabeel and Imrie, Fergus and van der Schaar, Mihaela},
booktitle={Advances in Neural Information Processing Systems},
year={2023}
}