metadata

license: apache-2.0
dataset_info:
  - config_name: arxiv
    features:
      - name: text
        dtype: string
    splits:
      - name: forget
        num_bytes: 22127152
        num_examples: 500
      - name: approximate
        num_bytes: 371246809
        num_examples: 6155
      - name: retain
        num_bytes: 84373706
        num_examples: 2000
    download_size: 216767075
    dataset_size: 477747667
  - config_name: general
    features:
      - name: text
        dtype: string
    splits:
      - name: evaluation
        num_bytes: 4628036
        num_examples: 1000
      - name: retain
        num_bytes: 24472399
        num_examples: 5000
    download_size: 17206310
    dataset_size: 29100435
  - config_name: github
    features:
      - name: text
        dtype: string
    splits:
      - name: forget
        num_bytes: 14069535
        num_examples: 2000
      - name: approximate
        num_bytes: 82904771
        num_examples: 15815
      - name: retain
        num_bytes: 28749659
        num_examples: 4000
    download_size: 43282163
    dataset_size: 125723965
configs:
  - config_name: arxiv
    data_files:
      - split: forget
        path: arxiv/forget-*
      - split: approximate
        path: arxiv/approximate-*
      - split: retain
        path: arxiv/retain-*
  - config_name: general
    data_files:
      - split: evaluation
        path: general/evaluation-*
      - split: retain
        path: general/retain-*
  - config_name: github
    data_files:
      - split: forget
        path: github/forget-*
      - split: approximate
        path: github/approximate-*
      - split: retain
        path: github/retain-*

📖 unlearn_dataset

The unlearn_dataset serves as a benchmark for evaluating unlearning methodologies in pre-trained large language models across diverse domains, including arXiv, GitHub.

🔍 Loading the datasets

To load the dataset:

from datasets import load_dataset

dataset = load_dataset("llmunlearn/unlearn_dataset", name="arxiv", split="forget")

Available configuration names and corresponding splits:
- arxiv: forget, approximate, retain
- github: forget, approximate, retain
- general: evaluation, retain

🛠️ Codebase

For evaluating unlearning methods on our datasets, visit our GitHub repository.

⭐ Citing our Work

If you find our codebase or dataset useful, please consider citing our paper:

@article{yao2024machine,
  title={Machine Unlearning of Pre-trained Large Language Models},
  author={Yao, Jin and Chien, Eli and Du, Minxin and Niu, Xinyao and Wang, Tianhao and Cheng, Zezhou and Yue, Xiang},
  journal={arXiv preprint arXiv:2402.15159},
  year={2024}
}