eval-unlearn-packages

Technique and metric implementation packages for eval-unlearn — a benchmarking framework for evaluating concept-unlearning techniques in text-to-image diffusion models.

eval-unlearn installs from PyPI as a lightweight core package; this repository hosts the standalone, installable implementations it delegates to for each unlearning technique and adversarial-attack metric, so users only pull in what they actually need:

• Techniques: esd, mace, uce, ssd, ca, cogfd, trasce, safree, advunlearn, concept-steerers, saeuron
• Adversarial-attack metrics: p4d, mma_diff, RING_A_BELL, Q16

Links

• Documentation: https://eval-unlearn.readthedocs.io/en/latest/getting-started/
• Leaderboard & interactive scoring tool: https://huggingface.co/spaces/REAL-Lab-Imperial/eval-unlearn
• Source code: https://github.com/REAL-Lab-Imperial/Eval-Unlearn
• PyPI package: https://pypi.org/project/eval-unlearn/

Usage

Some packages bundle large weight files tracked via Git LFS, so install by cloning rather than pip install git+...:

git clone https://huggingface.co/datasets/REAL-Lab-Imperial/eval-unlearn-packages
cd eval-unlearn-packages
git lfs pull

pip install -e esd/   # install only the packages you need

Full installation instructions, per-technique options, and metric configuration are covered in the Getting Started guide.

License

MIT