supreme-unlearning 0.1.0

A registry-based, multi-GPU framework for reproducible image-unlearning evaluation.

⚡ SUPREME - Standardised Unlearning Platform for REproducible Method Evaluation

🔬 Tech Stack
Core:
Accelerators:
Distributed & precision:

🛠️ Tooling
Experiment tracking:
Environment:
Debug & profile:
Code quality:

📄 Publication

📦 Repository

---

📖 Overview

SUPREME is an open-source framework for evaluating machine unlearning methods on image classification tasks.

What is machine unlearning? Given a model that was trained on some data, machine unlearning removes the influence of a chosen subset of that data (a class, a sub-class, or a random sample of examples) without retraining the model from scratch. Doing this well is hard: a good unlearned model should behave as if it had never seen the forgotten data, while still classifying everything else accurately. Many methods have been proposed, and they need a fair, repeatable way to be compared.

What SUPREME does. It runs the same three-stage pipeline end to end for any registered combination of dataset, model, unlearning method, and evaluation metric:

1. Train a baseline model on the full dataset.
2. Unlearn the chosen subset using the selected unlearning method.
3. Evaluate the unlearned model against a from-scratch retrained baseline (trained only on the data that was kept), using a configurable set of metrics that cover forgetting, utility, privacy, behavioural/parametric equivalence, and efficiency.

It ships 5 datasets, 2 model architectures, 2 baselines, 9 unlearning methods, and 9 selectable evaluation metrics (plus loss, reported automatically alongside accuracy), all selectable through command-line flags.

What makes SUPREME different:

• Reproducible. Recent work has shown that single-seed unlearning results can misrepresent a method's true behaviour. SUPREME runs the same experiment under multiple seeds, independently for the training, unlearning, and evaluation stages, so you measure distributions, not point estimates. The number of seeds at each stage is configurable per run.
• Multi-GPU and multi-precision. Built on PyTorch and Lightning Fabric. Distribution (DDP, FSDP, DeepSpeed ZeRO 1/2/3) applies to all three stages, with mixed-precision (fp16 / bf16) and NVIDIA / Apple Silicon / CPU back-ends.
• Registry-based extensibility. Add a dataset, model, unlearning method, or metric by implementing a small interface and registering its module path, with no framework changes required (see docs/extending.md).
• Efficient. When several experiments share the same training configuration, the model is trained once and reused across them, guarded by a file lock so parallel SLURM jobs and concurrent local runs stay consistent.

For the formal pipeline algorithm and mathematical notation (seed formulas, set definitions, operation signatures), see supreme/README.md and docs/notation.md.

---

🗃️ Available Components

Registry-based components are user-extensible - implement the relevant interface and register the module path, either in-tree or from your own package (runtime API or packaging entry points, no edits to SUPREME). See docs/extending.md. The components provided via Lightning Fabric cover the supported hardware and execution configurations.

Registry-based (user-extensible)

Component	Available implementations
Datasets	CIFAR-10, CIFAR-20, CIFAR-100, PinsFaceRecognition, Caltech-101
Models	ResNet18, Vision Transformer (ViT)
Baselines	Retrain, Original
Unlearning methods	Fine-Tuning (FT), Bad Teacher (BadT), Random Labels (RL), UNSIR, SSD, LFSSD, ASSD, SCRUB, JIT
Evaluation metrics	Accuracy, Loss/Error, ZRF, Activation Distance, JS-Divergence, Layer-wise Distance, Membership Inference Attack, Completeness, Resource Consumption, Time
Unlearning scenarios	Full-class, Subclass, Random sample

Provided via Lightning Fabric

Component	Available implementations
Accelerators	CPU, CUDA, MPS, TPU
Precision modes	64-true, 32-true, 16-mixed, bf16-mixed, 16-true, bf16-true, transformer-engine, transformer-engine-float16 (FP8), nf4, nf4-dq, fp4, fp4-dq, int8, int8-training
Distributed strategies	DDP, FSDP, DeepSpeed (ZeRO Stage 1/2/3)
Loggers	Weights & Biases, TensorBoard, CSV

---

⚡ Quickstart

# 1. Clone
git clone https://github.com/pedroandreou/supreme-unlearning.git
cd supreme-unlearning

# 2. Set up environment - the Makefile is the entry point for local dev: it creates
#    the venv (named `unlearning` by default; override with VENV=<name>), installs the
#    pinned deps + SUPREME (editable), and enables the git hook. (Prompts if it
#    already exists; pass ON_EXISTING=reuse to skip.)
make cuda                  # NVIDIA GPU (Linux / WSL2).  Apple Silicon / CPU: `make mps`
source unlearning/bin/activate

# 3. Configure W&B + HF tokens
cp .env.example .env
# edit .env with your WANDB_KEY, WANDB_USERNAME and HUGGING_FACE_HUB_TOKEN

# 4. Smoke test - one seed, one method, one dataset
bash supreme/run_local.sh \
  --gpu 0 --models ViT --training-seeds 260 \
  --methods retrain,finetune,ssd \
  --strategies random_ --datasets Cifar10 \
  --forget-percs 0.01

Full environment setup (Docker Dev Container, MPS prerequisites, etc.) is documented in docs/environment_setup.md. The Docker image is NVIDIA-only (Linux / WSL2); macOS users follow the virtual-env path above.

---

🧪 Running Experiments

The pipeline runs train → unlearn → evaluate automatically. Re-running is safe: per-stage outputs (training checkpoints, unlearning checkpoints, already-logged W&B results) are detected and skipped.

Local (workstation, GPU server, interactive cluster node)

# All 10 seeds, all methods, all datasets - defaults
bash supreme/run_local.sh --gpu 0

# Filter the sweep
bash supreme/run_local.sh \
  --gpu 0,1 \
  --models ViT \
  --training-seeds 260,261,262 \
  --methods retrain,finetune,bad_teacher,ssd \
  --strategies fullclass,random_ \
  --datasets PinsFaceRecognition

Flag	Description	Default
--gpu	GPU ID(s) - 0 single, 0,1,2,3 multi-GPU	0
--models	ResNet18, ViT	both
--training-seeds	Comma-separated training seeds (outer loop, I).	260–269
--unlearning-seeds	Space-separated indices for J (e.g. "0 1 2" for J=3)	"0" (matched)
--evaluation-seeds	Space-separated indices for K	"0" (matched)
--methods	Unlearning methods to run	all 11 (2 baselines + 9 methods)
--strategies	fullclass, subclass, random_	all
--datasets	Datasets to use	all 5
--forget-percs	Forget % for random_ strategy	0.001–0.10

SLURM (HPC, login node)

# Preview the grid (no submission)
./supreme/run_slurm.sh --dry-run

# Submit all experiments, max 12 concurrent jobs
./supreme/run_slurm.sh --max-concurrent 12

# Subset
./supreme/run_slurm.sh \
  --datasets Cifar10,Cifar20 \
  --models ViT \
  --training-seeds 260,261,262

# Multi-GPU DDP per job
./supreme/run_slurm.sh --gpus 4

Each submitted job runs one (seed, dataset, model) cell independently; cells run in parallel across the cluster. Distributed-strategy selection (DDP / FSDP / DeepSpeed) is documented in docs/implementation_notes.md → Distributed Strategies.

---

🔁 Reproducing the paper

Reproducing the paper's numbers is a two-step process: run the experiment grid on Pins Face Recognition (both architectures, both scenarios, all 10 seeds) and then render the three paper LaTeX tables from the W&B-logged results using supreme/utils/wandb_utils/results_analysis/pins_paper_tables.ipynb. The exact command, the table-rendering workflow, and the troubleshooting notes are documented in docs/reproducing_the_paper.md. For a runnable, step-by-step walkthrough (install → smoke test → full grid → tables → extending), see the notebook notebooks/reproduce_experiments.ipynb.

---

➕ Extending SUPREME

SUPREME is pip-installable (pip install supreme-unlearning, imported as supreme) and reusable as a library. Register your own components from your own package - no edits to framework code:

import supreme

supreme.register_unlearning_method("mymethod", "my_pkg.mymethod")
supreme.register_model("MyNet", "my_pkg.models:MyNet")
supreme.register_dataset("MyDS", "my_pkg.data:MyDS",
                         root="/data/myds", class_dict={"cat": 0, "dog": 1})

supreme.run_unlearning(["-method", "mymethod", "-net", "MyNet",
                        "-dataset", "MyDS", "-seed", "260"])

A runnable, end-to-end walkthrough from an external user's point of view - pip install supreme-unlearning then register your own method/metric/model/dataset - is in the notebook notebooks/custom_components.ipynb.

An installed plugin package can equivalently provide components via packaging entry points (supreme.models, supreme.unlearning_methods, supreme.metrics, supreme.plugins). The public API is supreme.register_*, supreme.run_training, supreme.run_unlearning, and supreme.project_config; everything under supreme.utils.* is internal.

Adding a dataset, model, method, or metric follows a consistent register-and-implement pattern. Walkthroughs and Fabric-integration rules live in docs/extending.md:

What to add	Walkthrough
New dataset	docs/extending.md → Adding a new dataset
New model	docs/extending.md → Adding a new model
New unlearning method	docs/extending.md → Adding a new unlearning method
New evaluation metric	docs/extending.md → Adding a new evaluation metric

---

🤝 Contributing

Contributions are welcome - bug reports, new components, and documentation alike.

• Found a bug or want a feature? Open an issue - the bug-report and feature-request templates appear automatically at New issue → choose a template.
• Adding a dataset, model, method, or metric? Most components register from your own package with no framework edits - see docs/extending.md. You can ship it as a pip-installable plugin or upstream it via a pull request.
• Opening a pull request? Run make style then make quality (the same ruff lint + format checks CI runs), and follow the PR template. Full workflow in the contributing guide.
• Share your method and results in community/ and add a row to the leaderboard.

CI (.github/workflows/ci.yml) lints, format-checks, and validates the package build on every push and PR. A version tag like v0.1.0 triggers .github/workflows/publish.yml to build and publish the release to PyPI (a manual run targets TestPyPI as a dry-run), and .github/workflows/docker.yml builds the CUDA image to GHCR. Notable changes per release are tracked in CHANGELOG.md.

---

📚 Documentation

Document	Covers
docs/contributing.md	How to report issues, add components, and open a pull request
CHANGELOG.md	Notable changes per release (Keep a Changelog / SemVer)
community/	Community-contributed methods, templates, and the results leaderboard
docs/notation.md	Symbol glossary - seeds, datasets, models, indices, counts
supreme/README.md	Formal algorithm specification (matched and decoupled protocols)
docs/environment_setup.md	Virtual-env and Docker Dev Container setup, .env template, prerequisites
docs/reproducing_the_paper.md	Single command for the paper's experiment grid plus the W&B-export-to-LaTeX-tables workflow
docs/script_arguments.md	Full argument reference for train_main.py and unlearn_main.py
docs/extending.md	How to add new datasets, models, methods, and metrics
docs/tooling.md	Debugger, profiler, Fabric callbacks, process tracker, split export, W&B exporter
docs/wandb_integration.md	W&B runtime behaviour: rank-0 logging, offline mode, sync workflow, metric synchronisation
docs/wandb_fields.md	Paper-to-W&B metric mapping and per-metric field paths
docs/implementation_notes.md	Distributed strategies, gradient handling, batch-size scaling, memory, known limitations
docs/adding_pinsfacerecognition.md	Manual Kaggle download for the Pins Face Recognition dataset
docs/future_work.md	Planned extensions

---

📝 Citing this work

@misc{supreme2026,
  title  = {SUPREME: Standardised Unlearning Platform for REproducible Method Evaluation},
  author = {Petros Andreou, Jamie Lanyon, Axel Finke, Georgina Cosma},
  year   = {2026},
  eprint = {2606.00380},
  archivePrefix = {arXiv},
  primaryClass = {cs.LG},
  url    = {https://arxiv.org/abs/2606.00380}
}

This work was conducted at Loughborough University.

---

📄 License

Released under the MIT License.

---

⭐ Star History