open-unlearning
A unified, easily extensible repository for LLM unlearning benchmarks (TOFU, MUSE) - enabling new evaluations, methods, and tasks.
Stars: 130
OpenUnlearning is an easily extensible framework that unifies LLM unlearning evaluation benchmarks. It provides efficient implementations of TOFU and MUSE unlearning benchmarks, supporting 5 unlearning methods, 3+ datasets, 6+ evaluation metrics, and 7+ LLMs. Users can easily extend the framework to incorporate more variants, collaborate by adding new benchmarks, unlearning methods, datasets, and evaluation metrics, and drive progress in the field.
README:
We provide efficient and streamlined implementations of the TOFU, MUSE unlearning benchmarks while supporting 5 unlearning methods, 3+ datasets, 6+ evaluation metrics, and 7+ LLMs. Each of these can be easily extended to incorporate more variants.
We invite the LLM unlearning community to collaborate by adding new benchmarks, unlearning methods, datasets and evaluation metrics here to expand OpenUnlearning's features, gain feedback from wider usage and drive progress in the field.
⚠️ Notice (Updated: February 27, 2025)
This repository replaces the original TOFU codebase, which can be found atgithub.com/locuslab/tofuand isn't maintained anymore.
We provide several variants for each of the components in the unlearning pipeline.
| Component | Available Options |
|---|---|
| Benchmarks | TOFU, MUSE |
| Unlearning Methods | GradAscent, GradDiff, NPO, SimNPO, DPO |
| Evaluation Metrics | Verbatim Probability, Verbatim ROUGE, QA-ROUGE, MIA Attacks, TruthRatio, Model Utility |
| Datasets | MUSE-News (BBC), MUSE-Books (Harry Potter), TOFU (different splits) |
| Model Families | TOFU: LLaMA-3.2, LLaMA-3.1, LLaMA-2; MUSE: LLaMA-2, ICLM; Additional: Phi-3.5, Phi-1.5, Gemma |
- 📖 Overview
- 🗃️ Available Components
- ⚡ Quickstart
- 🔄 Updated TOFU benchmark
- 🧪 Running Experiments
- ➕ How to Add New Components
- 📚 Further Documentation
- 🔗 Support & Contributors
- 📝 Citing this work
- 🤝 Acknowledgements
- 📄 License
conda create -n unlearning python=3.11
conda activate unlearning
pip install .
pip install --no-build-isolation flash-attn==2.6.3Download the log files containing metric results from the models used in the supported benchmarks (including the retain model logs used to compare the unlearned models against).
python setup_data.py # populates saves/eval with evaluation results of the uploaded modelsWe've updated Open-Unlearning's TOFU benchmark target models to use a wider variety of newer architectures with sizes varying from 1B to 8B. These include LLaMA 3.2 1B, LLaMA 3.2 3B, LLaMA 3.1 8B, and the original LLaMA-2 7B from the old version of TOFU.
For each architecture, we have finetuned with four different splits of the TOFU datasets: full, retain90, retain95, retain99, for a total of 16 finetuned models. The first serves as the target (base model for unlearning) and the rest are retain models used to measure performance against for each forget split. These models are on HuggingFace and the paths to these models can be set in the experimental configs or in command-line overrides.
We provide an easily configurable interface for running evaluations by leveraging Hydra configs. For a more detailed documentation of aspects like running experiments, commonly overriden arguments, interfacing with configurations, distributed training and simple finetuning of models, refer docs/experiments.md.
An example command for launching an unlearning process with GradAscent on the TOFU forget10 split:
python src/train.py --config-name=unlearn.yaml experiment=unlearn/tofu/default \
forget_split=forget10 retain_split=retain90 trainer=GradAscent task_name=SAMPLE_UNLEARN-
experiment- Path to the Hydra config fileconfigs/experiment/unlearn/muse/default.yamlwith default experimental settings for TOFU unlearning, e.g. train dataset, eval benchmark details, model paths etc.. -
forget_split/retain_split- Sets the forget and retain dataset splits. -
trainer- Loadconfigs/trainer/GradAscent.yamland override the unlearning method with the handler (see config) implemented insrc/trainer/unlearn/grad_ascent.py.
An example command for launching a TOFU evaluation process on forget10 split:
python src/eval.py --config-name=eval.yaml experiment=eval/tofu/default \
model=Llama-3.2-1B-Instruct \
model.model_args.pretrained_model_name_or_path=open-unlearning/tofu_Llama-3.2-1B-Instruct_full \
task_name=SAMPLE_EVAL-
experiment- Path to the evaluation configurationconfigs/experiment/eval/tofu/default.yaml. -
model- Sets up the model and tokenizer configs for theLlama-3.2-1B-Instructmodel. -
model.model_args.pretrained_model_name_or_path- Overrides the default experiment config to evaluate a model from a HuggingFace ID (can use a local model checkpoint path as well).
For more details about creating and running evaluations, refer docs/evaluation.md.
The scripts below execute standard baseline unlearning experiments on the TOFU and MUSE datasets, evaluated using their corresponding benchmarks. The expected results for these are in docs/results.md.
bash scripts/tofu_unlearn.sh
bash scripts/muse_unlearn.shAdding a new component (trainer, evaluation metric, benchmark, model, or dataset) requires defining a new class, registering it, and creating a configuration file. Learn more about adding new components in docs/components.md.
Please feel free to raise a pull request for any new features after setting up the environment in development mode.
pip install .[dev]For more in-depth information on specific aspects of the framework, refer to the following documents:
| Documentation | Contains |
|---|---|
docs/components.md |
Instructions on how to add new components such as trainers, benchmarks, metrics, models, datasets, etc. |
docs/evaluation.md |
Detailed instructions on creating and running evaluation metrics and benchmarks. |
docs/experiments.md |
Guide on running experiments in various configurations and settings, including distributed training, fine-tuning, and overriding arguments. |
docs/hydra.md |
Explanation of the Hydra features used in configuration management for experiments. |
docs/results.md |
Reference results from various unlearning methods run using this framework on TOFU and MUSE benchmarks. |
Developed and maintained by Vineeth Dorna (@Dornavineeth) and Anmol Mekala (@molereddy).
If you encounter any issues or have questions, feel free to raise an issue in the repository 🛠️.
If you use OpenUnlearning in your research, please cite:
@misc{openunlearning2025,
title={OpenUnlearning: A Unified Framework for LLM Unlearning Benchmarks},
author={Dorna, Vineeth and Mekala, Anmol and Zhao, Wenlong and McCallum, Andrew and Kolter, J Zico and Maini, Pratyush},
year={2025},
howpublished={\url{https://github.com/locuslab/open-unlearning}},
note={Accessed: February 27, 2025}
}
@inproceedings{maini2024tofu,
title={TOFU: A Task of Fictitious Unlearning for LLMs},
author={Maini, Pratyush and Feng, Zhili and Schwarzschild, Avi and Lipton, Zachary Chase and Kolter, J Zico},
booktitle={First Conference on Language Modeling},
year={2024}
}Expand for bibtex to cite other benchmarks used from OpenUnlearning
@article{shi2024muse,
title={Muse: Machine unlearning six-way evaluation for language models},
author={Shi, Weijia and Lee, Jaechan and Huang, Yangsibo and Malladi, Sadhika and Zhao, Jieyu and Holtzman, Ari and Liu, Daogao and Zettlemoyer, Luke and Smith, Noah A and Zhang, Chiyuan},
journal={arXiv preprint arXiv:2407.06460},
year={2024}
}- This repo is inspired from LLaMA-Factory.
- The TOFU and MUSE benchmarks served as the foundation for our re-implementation.
This project is licensed under the MIT License. See the LICENSE file for details.
For Tasks:
Click tags to check more tools for each tasksFor Jobs:
Alternative AI tools for open-unlearning
Similar Open Source Tools
open-unlearning
OpenUnlearning is an easily extensible framework that unifies LLM unlearning evaluation benchmarks. It provides efficient implementations of TOFU and MUSE unlearning benchmarks, supporting 5 unlearning methods, 3+ datasets, 6+ evaluation metrics, and 7+ LLMs. Users can easily extend the framework to incorporate more variants, collaborate by adding new benchmarks, unlearning methods, datasets, and evaluation metrics, and drive progress in the field.
wanda
Official PyTorch implementation of Wanda (Pruning by Weights and Activations), a simple and effective pruning approach for large language models. The pruning approach removes weights on a per-output basis, by the product of weight magnitudes and input activation norms. The repository provides support for various features such as LLaMA-2, ablation study on OBS weight update, zero-shot evaluation, and speedup evaluation. Users can replicate main results from the paper using provided bash commands. The tool aims to enhance the efficiency and performance of language models through structured and unstructured sparsity techniques.
FlexFlow
FlexFlow Serve is an open-source compiler and distributed system for **low latency**, **high performance** LLM serving. FlexFlow Serve outperforms existing systems by 1.3-2.0x for single-node, multi-GPU inference and by 1.4-2.4x for multi-node, multi-GPU inference.
LLM-Pruner
LLM-Pruner is a tool for structural pruning of large language models, allowing task-agnostic compression while retaining multi-task solving ability. It supports automatic structural pruning of various LLMs with minimal human effort. The tool is efficient, requiring only 3 minutes for pruning and 3 hours for post-training. Supported LLMs include Llama-3.1, Llama-3, Llama-2, LLaMA, BLOOM, Vicuna, and Baichuan. Updates include support for new LLMs like GQA and BLOOM, as well as fine-tuning results achieving high accuracy. The tool provides step-by-step instructions for pruning, post-training, and evaluation, along with a Gradio interface for text generation. Limitations include issues with generating repetitive or nonsensical tokens in compressed models and manual operations for certain models.
AgentBench
AgentBench is a benchmark designed to evaluate Large Language Models (LLMs) as autonomous agents in various environments. It includes 8 distinct environments such as Operating System, Database, Knowledge Graph, Digital Card Game, and Lateral Thinking Puzzles. The tool provides a comprehensive evaluation of LLMs' ability to operate as agents by offering Dev and Test sets for each environment. Users can quickly start using the tool by following the provided steps, configuring the agent, starting task servers, and assigning tasks. AgentBench aims to bridge the gap between LLMs' proficiency as agents and their practical usability.
PDEBench
PDEBench provides a diverse and comprehensive set of benchmarks for scientific machine learning, including challenging and realistic physical problems. The repository consists of code for generating datasets, uploading and downloading datasets, training and evaluating machine learning models as baselines. It features a wide range of PDEs, realistic and difficult problems, ready-to-use datasets with various conditions and parameters. PDEBench aims for extensibility and invites participation from the SciML community to improve and extend the benchmark.
flashinfer
FlashInfer is a library for Language Languages Models that provides high-performance implementation of LLM GPU kernels such as FlashAttention, PageAttention and LoRA. FlashInfer focus on LLM serving and inference, and delivers state-the-art performance across diverse scenarios.
AdalFlow
AdalFlow is a library designed to help developers build and optimize Large Language Model (LLM) task pipelines. It follows a design pattern similar to PyTorch, offering a light, modular, and robust codebase. Named in honor of Ada Lovelace, AdalFlow aims to inspire more women to enter the AI field. The library is tailored for various GenAI applications like chatbots, translation, summarization, code generation, and autonomous agents, as well as classical NLP tasks such as text classification and named entity recognition. AdalFlow emphasizes modularity, robustness, and readability to support users in customizing and iterating code for their specific use cases.
evalverse
Evalverse is an open-source project designed to support Large Language Model (LLM) evaluation needs. It provides a standardized and user-friendly solution for processing and managing LLM evaluations, catering to AI research engineers and scientists. Evalverse supports various evaluation methods, insightful reports, and no-code evaluation processes. Users can access unified evaluation with submodules, request evaluations without code via Slack bot, and obtain comprehensive reports with scores, rankings, and visuals. The tool allows for easy comparison of scores across different models and swift addition of new evaluation tools.
xFasterTransformer
xFasterTransformer is an optimized solution for Large Language Models (LLMs) on the X86 platform, providing high performance and scalability for inference on mainstream LLM models. It offers C++ and Python APIs for easy integration, along with example codes and benchmark scripts. Users can prepare models in a different format, convert them, and use the APIs for tasks like encoding input prompts, generating token ids, and serving inference requests. The tool supports various data types and models, and can run in single or multi-rank modes using MPI. A web demo based on Gradio is available for popular LLM models like ChatGLM and Llama2. Benchmark scripts help evaluate model inference performance quickly, and MLServer enables serving with REST and gRPC interfaces.
basiclingua-LLM-Based-NLP
BasicLingua is a Python library that provides functionalities for linguistic tasks such as tokenization, stemming, lemmatization, and many others. It is based on the Gemini Language Model, which has demonstrated promising results in dealing with text data. BasicLingua can be used as an API or through a web demo. It is available under the MIT license and can be used in various projects.
qlib
Qlib is an open-source, AI-oriented quantitative investment platform that supports diverse machine learning modeling paradigms, including supervised learning, market dynamics modeling, and reinforcement learning. It covers the entire chain of quantitative investment, from alpha seeking to order execution. The platform empowers researchers to explore ideas and implement productions using AI technologies in quantitative investment. Qlib collaboratively solves key challenges in quantitative investment by releasing state-of-the-art research works in various paradigms. It provides a full ML pipeline for data processing, model training, and back-testing, enabling users to perform tasks such as forecasting market patterns, adapting to market dynamics, and modeling continuous investment decisions.
LMCache
LMCache is a serving engine extension designed to reduce time to first token (TTFT) and increase throughput, particularly in long-context scenarios. It stores key-value caches of reusable texts across different locations like GPU, CPU DRAM, and Local Disk, allowing the reuse of any text in any serving engine instance. By combining LMCache with vLLM, significant delay savings and GPU cycle reduction are achieved in various large language model (LLM) use cases, such as multi-round question answering and retrieval-augmented generation (RAG). LMCache provides integration with the latest vLLM version, offering both online serving and offline inference capabilities. It supports sharing key-value caches across multiple vLLM instances and aims to provide stable support for non-prefix key-value caches along with user and developer documentation.
notte
Notte is a web browser designed specifically for LLM agents, providing a language-first web navigation experience without the need for DOM/HTML parsing. It transforms websites into structured, navigable maps described in natural language, enabling users to interact with the web using natural language commands. By simplifying browser complexity, Notte allows LLM policies to focus on conversational reasoning and planning, reducing token usage, costs, and latency. The tool supports various language model providers and offers a reinforcement learning style action space and controls for full navigation control.
rag-chatbot
The RAG ChatBot project combines Lama.cpp, Chroma, and Streamlit to build a Conversation-aware Chatbot and a Retrieval-augmented generation (RAG) ChatBot. The RAG Chatbot works by taking a collection of Markdown files as input and provides answers based on the context provided by those files. It utilizes a Memory Builder component to load Markdown pages, divide them into sections, calculate embeddings, and save them in an embedding database. The chatbot retrieves relevant sections from the database, rewrites questions for optimal retrieval, and generates answers using a local language model. It also remembers previous interactions for more accurate responses. Various strategies are implemented to deal with context overflows, including creating and refining context, hierarchical summarization, and async hierarchical summarization.
LLamaSharp
LLamaSharp is a cross-platform library to run 🦙LLaMA/LLaVA model (and others) on your local device. Based on llama.cpp, inference with LLamaSharp is efficient on both CPU and GPU. With the higher-level APIs and RAG support, it's convenient to deploy LLM (Large Language Model) in your application with LLamaSharp.
For similar tasks
open-unlearning
OpenUnlearning is an easily extensible framework that unifies LLM unlearning evaluation benchmarks. It provides efficient implementations of TOFU and MUSE unlearning benchmarks, supporting 5 unlearning methods, 3+ datasets, 6+ evaluation metrics, and 7+ LLMs. Users can easily extend the framework to incorporate more variants, collaborate by adding new benchmarks, unlearning methods, datasets, and evaluation metrics, and drive progress in the field.
labelbox-python
Labelbox is a data-centric AI platform for enterprises to develop, optimize, and use AI to solve problems and power new products and services. Enterprises use Labelbox to curate data, generate high-quality human feedback data for computer vision and LLMs, evaluate model performance, and automate tasks by combining AI and human-centric workflows. The academic & research community uses Labelbox for cutting-edge AI research.
promptfoo
Promptfoo is a tool for testing and evaluating LLM output quality. With promptfoo, you can build reliable prompts, models, and RAGs with benchmarks specific to your use-case, speed up evaluations with caching, concurrency, and live reloading, score outputs automatically by defining metrics, use as a CLI, library, or in CI/CD, and use OpenAI, Anthropic, Azure, Google, HuggingFace, open-source models like Llama, or integrate custom API providers for any LLM API.
vespa
Vespa is a platform that performs operations such as selecting a subset of data in a large corpus, evaluating machine-learned models over the selected data, organizing and aggregating it, and returning it, typically in less than 100 milliseconds, all while the data corpus is continuously changing. It has been in development for many years and is used on a number of large internet services and apps which serve hundreds of thousands of queries from Vespa per second.
python-aiplatform
The Vertex AI SDK for Python is a library that provides a convenient way to use the Vertex AI API. It offers a high-level interface for creating and managing Vertex AI resources, such as datasets, models, and endpoints. The SDK also provides support for training and deploying custom models, as well as using AutoML models. With the Vertex AI SDK for Python, you can quickly and easily build and deploy machine learning models on Vertex AI.
ScandEval
ScandEval is a framework for evaluating pretrained language models on mono- or multilingual language tasks. It provides a unified interface for benchmarking models on a variety of tasks, including sentiment analysis, question answering, and machine translation. ScandEval is designed to be easy to use and extensible, making it a valuable tool for researchers and practitioners alike.
opencompass
OpenCompass is a one-stop platform for large model evaluation, aiming to provide a fair, open, and reproducible benchmark for large model evaluation. Its main features include: * Comprehensive support for models and datasets: Pre-support for 20+ HuggingFace and API models, a model evaluation scheme of 70+ datasets with about 400,000 questions, comprehensively evaluating the capabilities of the models in five dimensions. * Efficient distributed evaluation: One line command to implement task division and distributed evaluation, completing the full evaluation of billion-scale models in just a few hours. * Diversified evaluation paradigms: Support for zero-shot, few-shot, and chain-of-thought evaluations, combined with standard or dialogue-type prompt templates, to easily stimulate the maximum performance of various models. * Modular design with high extensibility: Want to add new models or datasets, customize an advanced task division strategy, or even support a new cluster management system? Everything about OpenCompass can be easily expanded! * Experiment management and reporting mechanism: Use config files to fully record each experiment, and support real-time reporting of results.
flower
Flower is a framework for building federated learning systems. It is designed to be customizable, extensible, framework-agnostic, and understandable. Flower can be used with any machine learning framework, for example, PyTorch, TensorFlow, Hugging Face Transformers, PyTorch Lightning, scikit-learn, JAX, TFLite, MONAI, fastai, MLX, XGBoost, Pandas for federated analytics, or even raw NumPy for users who enjoy computing gradients by hand.
For similar jobs
sweep
Sweep is an AI junior developer that turns bugs and feature requests into code changes. It automatically handles developer experience improvements like adding type hints and improving test coverage.
teams-ai
The Teams AI Library is a software development kit (SDK) that helps developers create bots that can interact with Teams and Microsoft 365 applications. It is built on top of the Bot Framework SDK and simplifies the process of developing bots that interact with Teams' artificial intelligence capabilities. The SDK is available for JavaScript/TypeScript, .NET, and Python.
ai-guide
This guide is dedicated to Large Language Models (LLMs) that you can run on your home computer. It assumes your PC is a lower-end, non-gaming setup.
classifai
Supercharge WordPress Content Workflows and Engagement with Artificial Intelligence. Tap into leading cloud-based services like OpenAI, Microsoft Azure AI, Google Gemini and IBM Watson to augment your WordPress-powered websites. Publish content faster while improving SEO performance and increasing audience engagement. ClassifAI integrates Artificial Intelligence and Machine Learning technologies to lighten your workload and eliminate tedious tasks, giving you more time to create original content that matters.
chatbot-ui
Chatbot UI is an open-source AI chat app that allows users to create and deploy their own AI chatbots. It is easy to use and can be customized to fit any need. Chatbot UI is perfect for businesses, developers, and anyone who wants to create a chatbot.
BricksLLM
BricksLLM is a cloud native AI gateway written in Go. Currently, it provides native support for OpenAI, Anthropic, Azure OpenAI and vLLM. BricksLLM aims to provide enterprise level infrastructure that can power any LLM production use cases. Here are some use cases for BricksLLM: * Set LLM usage limits for users on different pricing tiers * Track LLM usage on a per user and per organization basis * Block or redact requests containing PIIs * Improve LLM reliability with failovers, retries and caching * Distribute API keys with rate limits and cost limits for internal development/production use cases * Distribute API keys with rate limits and cost limits for students
uAgents
uAgents is a Python library developed by Fetch.ai that allows for the creation of autonomous AI agents. These agents can perform various tasks on a schedule or take action on various events. uAgents are easy to create and manage, and they are connected to a fast-growing network of other uAgents. They are also secure, with cryptographically secured messages and wallets.
griptape
Griptape is a modular Python framework for building AI-powered applications that securely connect to your enterprise data and APIs. It offers developers the ability to maintain control and flexibility at every step. Griptape's core components include Structures (Agents, Pipelines, and Workflows), Tasks, Tools, Memory (Conversation Memory, Task Memory, and Meta Memory), Drivers (Prompt and Embedding Drivers, Vector Store Drivers, Image Generation Drivers, Image Query Drivers, SQL Drivers, Web Scraper Drivers, and Conversation Memory Drivers), Engines (Query Engines, Extraction Engines, Summary Engines, Image Generation Engines, and Image Query Engines), and additional components (Rulesets, Loaders, Artifacts, Chunkers, and Tokenizers). Griptape enables developers to create AI-powered applications with ease and efficiency.