https://github.com/jeakwon/ai-engram

Surgical unlearning of classes, concepts, and facts in any PyTorch model
https://github.com/jeakwon/ai-engram

deep-learning machine-unlearning model-editing pytorch

Last synced: 10 days ago
JSON representation

Surgical unlearning of classes, concepts, and facts in any PyTorch model

• Host: GitHub
• URL: https://github.com/jeakwon/ai-engram
• Owner: jeakwon
• License: mit
• Created: 2026-02-09T17:56:01.000Z (4 months ago)
• Default Branch: main
• Last Pushed: 2026-05-24T19:25:49.000Z (10 days ago)
• Last Synced: 2026-05-24T19:28:06.413Z (10 days ago)
• Topics: deep-learning, machine-unlearning, model-editing, pytorch
• Language: Python
• Homepage: https://jeakwon.github.io/ai-engram/
• Size: 59.6 KB
• Stars: 0
• Watchers: 0
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • Changelog: CHANGELOG.md
  • License: LICENSE

Awesome Lists containing this project

README

          
# ai-engram

**Surgical unlearning of classes, concepts, and facts in any PyTorch model.**

[![PyPI version](https://img.shields.io/pypi/v/ai-engram.svg)](https://pypi.org/project/ai-engram/)

[![Python versions](https://img.shields.io/pypi/pyversions/ai-engram.svg)](https://pypi.org/project/ai-engram/)

[![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](LICENSE)

[![CI](https://github.com/jeakwon/ai-engram/actions/workflows/ci.yml/badge.svg)](https://github.com/jeakwon/ai-engram/actions/workflows/ci.yml)

[![Docs](https://img.shields.io/badge/docs-online-success.svg)](https://jeakwon.github.io/ai-engram/)

`ai-engram` identifies the directions in a network's weights that store

specific information — a class, a concept, a fact — and edits them out

in closed form, with no gradient descent. It works on `Linear`,

`Conv1d`, and `Conv2d` layers (grouped and depthwise convolutions

included) and has been verified end-to-end against 24 pretrained

vision, language, and diffusion models.

## Installation

```bash

pip install ai-engram

```

## Quickstart

```python

import torch

import torch.nn as nn

from torch.utils.data import DataLoader, TensorDataset

from ai_engram import EditorConfig, EngramEditor

model = nn.Sequential(nn.Linear(8, 16), nn.ReLU(), nn.Linear(16, 4))

forget_loader = DataLoader(TensorDataset(torch.randn(32, 8)), batch_size=8)

total_loader = DataLoader(TensorDataset(torch.randn(128, 8)), batch_size=8)

editor = EngramEditor(model, EditorConfig(device="cpu"))

forget_cov = editor.collect_statistics(forget_loader)

total_cov = editor.collect_statistics(total_loader)

edited = editor.edit(forget_cov, total_cov)

```

`forget_cov` carries the second-moment statistics of the data you want

the model to forget; `total_cov` carries the statistics of the full

training distribution. The edit removes the projection of the weights

onto the forget subspace while preserving the rest.

This exact snippet is executed in CI on every commit and on every

release, so a `pip install` always produces a runnable library.

## Documentation

Full guides, API reference, and the 24-model compatibility report live

at .

## License

[MIT](LICENSE) — see also [CHANGELOG.md](CHANGELOG.md).