FiMODA: Final-Model-Only Data Attribution

This repository provides code for reproducing experiments in the NeurIPS 2025 paper "Final-Model-Only Data Attribution with a Unifying View of Gradient-Based Methods."

Installation

uv

We recommend using uv as the package manager. If needed, install uv via either:

curl -Ls https://astral.sh/uv/install.sh | sh

or using Homebrew:

brew install astral-sh/uv/uv

or using pip (use this for Windows):

pip install uv

FiMODA

First clone the repository:

git clone [email protected]:IBM/fimoda.git
cd fimoda

Once inside the fimoda directory (where this README.md is located), if using Linux or Mac, run:

uv venv --python 3.12
source .venv/bin/activate
uv pip install .

Or in Windows, run:

uv venv --python 3.12
.venv/bin/activate
uv pip install .

The code has been tested on Red Hat Enterprise Linux 9 and macOS (tabular subfolder only for the latter).

Organization

The code is organized into three subfolders under the fimoda folder, corresponding to three data modalities: tabular, text, and image. The code within each of these subfolders is mostly self-contained.

At least one GPU is strongly recommended for running the code in the text and image subfolders.

Citation and appreciation

If you find our work useful, please star the repository and cite our work as follows:

@inproceedings{
    wei2025fimoda,
    title={Final-Model-Only Data Attribution with a Unifying View of Gradient-Based Methods},
    author={Dennis Wei and Inkit Padhi and Soumya Ghosh and Amit Dhurandhar and Karthikeyan Natesan Ramamurthy and Maria Chang},
    booktitle={The Thirty-ninth Annual Conference on Neural Information Processing Systems},
    year={2025},
    url={https://openreview.net/forum?id=rccgEdFTlH}
}

License

This code is provided under the Apache 2.0 license.