Datasets:

ClaudioSavelli
/

TOFU

license: mit
language:
  - en
pretty_name: TOFU with author identity names
task_categories:
  - question-answering
tags:
  - unlearning
  - tofu
  - question-answering
dataset_info:
  config_name: full
  features:
    - name: question
      dtype: string
    - name: answer
      dtype: string
    - name: name
      dtype: string
  splits:
    - name: train
      num_bytes: 1070038
      num_examples: 4000
  download_size: 499434
  dataset_size: 1070038
configs:
  - config_name: full
    data_files:
      - split: train
        path: full/train-*

TOFU with author identity names

This is a lossless, order-preserving derivative of the full configuration of locuslab/TOFU. It adds one column, name, so that all question-answer pairs about the same fictitious author can be grouped as one deletion entity.

The source data was pinned at commit 324592d84ae4f482ac7249b9285c2ecdb53e3a68. The question and answer values and their original order were not changed.

Loading

from datasets import load_dataset

dataset = load_dataset("ClaudioSavelli/TOFU", "full", split="train")

The columns are:

question: the original TOFU question.
answer: the original TOFU answer.
name: the full fictitious-author name assigned to the row.

This repository currently contains only the complete 4,000-row source cohort. It does not yet define retain, forget, test, or sequential deletion-request splits.

Identity construction and validation

The upstream full data is ordered as 200 consecutive author blocks, with 20 question-answer pairs per author. Names were recovered from the text within each block, and block membership was used for rows that refer to the author only with a pronoun or shortened name.

Validation results:

4,000 rows and 200 distinct identity values;
exactly 20 rows for every identity value;
3,969 rows contain their assigned full name verbatim in the question or answer;
11 rows belonging to named authors omit the full name but are unambiguous from their surrounding source block;
one complete 20-row upstream block never states an author name.

The nameless source block is the Karachi-born Urban Fiction author at zero-based identity position 88. Its rows are preserved and assigned the explicit placeholder __UNKNOWN_TOFU_IDENTITY_088__. This is deliberately not a fabricated person name. Consumers should treat the placeholder as one entity and decide explicitly whether to retain, exclude, or repair that entity before experiments that require every entity to have a natural-language name.

License and citation

This derivative retains the upstream MIT license. Please cite the original TOFU 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}
}