A web platform to democratize Machine Unlearning. Quick demo of the Forgetsy web UI in action (respectively exploring the results of a benchmark and running a new unlearnign session):

It enables experimenting with unlearning in a no-code manner, enabling a broader public (such as policy-makers, philosofers, among others) to exlore the benefits and shortcomings of Machine Unlearning.

It is built fully on top of the Vision Unlearning library, and the benchmark exploration feature is compatible with any HuggingFace repository whose structure follows Appendix 2 of I-CARE paper.

Getting started

Common setup (exploring benchmarks: listing entities, running Result Templates)

This path needs no .env file and no secrets at all — forgety reads benchmark data from a public HuggingFace repository anonymously.

1. Clone this repository and vision-unlearning as sibling directories — docker-compose.yml's bind mounts require exactly this layout:

   parent/
   ├── forgety/            (this repository)
   └── vision-unlearning/
   

   (vision-unlearning is not installed as a Python package today; it's mounted directly from the sibling checkout. See the TODO comments in docker-compose.yml — publishing it to PyPI and dropping the mount is a separate, bigger task.)
2. From forgety/, run:

   docker compose up
   

3. Go to http://localhost:8501. Backend docs are at http://localhost:8001/docs.

This is also the starting point for running unlearning sessions (Slurm), below — that path is a strict superset of this one, needing real credentials on top of it.

Additional setup for running unlearning sessions (Slurm)

Running a new unlearning session (as opposed to exploring existing benchmark results) needs real credentials: a HuggingFace token, to upload the resulting model, and SSH/Slurm credentials, to submit the job to a cluster. Currently only a Slurm cluster is supported as the execution backend.

1. Copy .env.template to .env and fill in HF_TOKEN plus the SSH/Slurm variables (see the template for the exact keys expected).
2. The same .env content is needed in three locations, all git-ignored:
   • .env — used by docker compose (via env_file) for the services running locally.
   • infra/.env — needed because the HF credential must also be available to the code that runs on the cluster, to allow uploading the resulting model.
   • services/backend/app/.env — needed for the backend to connect to Slurm.
3. Server-side setup: a server must be configured to actually execute the job (currently only Slurm clusters are supported). Our setup was tested on the PPKE's ITK HPC cluster (Esztergom) in December 2025, with:
   • Tesla V100-PCIE-16GB with CUDA Version: 12.6
   • Slurm 23.11.4
   • RHEL/CentOS/Fedora 8.10 (Green Obsidian)
   • vision-unlearning 0.1.6 with Python 3.10

Testing

Tests run automatically on GitHub Actions for every push and pull request (.github/workflows/test.yml): mypy, pycodestyle, and the offline pytest suite (backend API tests, I-CARE route tests, and Streamlit UI tests via streamlit.testing.v1.AppTest — no browser or running backend needed).

To run them locally:

make test          # inside Docker (mypy + pycodestyle + offline pytest)

or on the host, with the dependencies from services/backend/requirements.txt, services/frontend/requirements.txt, libs/requirements.txt and libs/requirements.dev.txt installed, and a checkout of vision-unlearning as a sibling directory of this repository:

python -m pytest tests -m "not gpu and not integration"

Tests marked integration download real I-CARE data from HuggingFace; they are excluded from all default runs and executed weekly (or on demand) by .github/workflows/integration.yml.

Tech stack

• Automation and devops: Docker, makefile, mypi, PEP8, pytest, pypi, readthedocs, git, github (issues and task board)
• Frontend: Streamlit, python
• Backend: FastAPI, pydantic, jsonschema, python
• Infrastructure and cluster integration: SSH, bash, Slurm commands, apptainer

Architecture