TOFU

TOFU: Task of Fictitious Unlearning

The TOFU dataset serves as a benchmark for evaluating unlearning performance of large language models on realistic tasks. The dataset comprises question-answer pairs based on autobiographies of 200 different authors that do not exist and are completely fictitiously generated by the GPT-4 model. The goal of the task is to unlearn a fine-tuned model on various fractions of the forget set.

Quick Links

• Website: The landing page for TOFU
• arXiv Paper: Detailed information about the TOFU dataset and its significance in unlearning tasks.
• GitHub Repository: Access the source code, fine-tuning scripts, and additional resources for the TOFU dataset.
• Dataset on Hugging Face: Direct link to download the TOFU dataset.
• Leaderboard on Hugging Face Spaces: Current rankings and submissions for the TOFU dataset challenges.
• Summary on Twitter: A concise summary and key takeaways from the project.

Applicability

The dataset is in QA format, making it ideal for use with popular chat models such as Llama2, Mistral, or Qwen. However, it also works for any other large language model. The corresponding code base is written for the Llama2 chat, and Phi-1.5 models, but can be easily adapted to other models.

Loading the Dataset

To load the dataset, use the following code:

from datasets import load_dataset
dataset = load_dataset("locuslab/TOFU", "full")

Available forget sets are:

• forget01: Forgetting 1% of the original dataset, all entries correspond to a single author.
• forget05: Forgetting 5% of the original dataset, all entries correspond to a single author.
• forget10: Forgetting 10% of the original dataset, all entries correspond to a single author.

Retain sets corresponding to each forget set are also available, which can be used to train an Oracle model.

Codebase

The code for training the models and the availability of all fine-tuned models can be found at our GitHub repository.

Citing Our Work

If you find our codebase and dataset beneficial, please cite our work:

@misc{tofu2024,
      title={TOFU: A Task of Fictitious Unlearning for LLMs}, 
      author={Pratyush Maini and Zhili Feng and Avi Schwarzschild and Zachary C. Lipton and J. Zico Kolter},
      year={2024},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}