sungjuncho/TOFU-para

TOFU, typo-corrected, with 5 paraphrase versions of every QA pair

A drop-in copy of TOFU (Maini et al., 2024) with a handful of name typos corrected, plus five independently generated, fact-checked paraphrases of all 4000 question-answer pairs. Every TOFU subset has the same name, rows and row order as the original, so

from datasets import load_dataset
load_dataset("sungjuncho/TOFU-para", "forget10")         # corrected original forget10
load_dataset("sungjuncho/TOFU-para", "forget10_para1")   # paraphrase version 1 of the same rows

Built for studying whether unlearning methods remove knowledge or only the specific training samples: a model can be finetuned on the originals and some paraphrase versions, unlearned on the originals only, and probed with the versions it saw and the ones it did not. Models trained on this data: sungjuncho/Llama-3.1-8B-Instruct-TOFU-para.

Contents

Config	Rows	What it is
full, forget01/05/10, retain90/95/99	4000, 40/200/400, 3600/3800/3960	TOFU's training splits, corrected
<split>_para1 ... <split>_para5 (each of the 7 splits)	same as <split>	Paraphrase version k of <split>: row i is a rewording of row i of <split>
forget01/05/10_perturbed, retain_perturbed, holdout01/05/10, world_facts(_perturbed), real_authors(_perturbed)	as in TOFU	TOFU's evaluation subsets, corrected (their own paraphrased_* and perturbed_* fields are TOFU's, unchanged apart from the corrections)

The paraphrase files have only question and answer fields; both are rewritten. The paraphrases of a split are exactly the corresponding rows of full_para<k> (forget10 is rows 3600-3999 of full, as in TOFU), so the versions of one fact can be matched across files by row index.

Typo corrections

Five rows of TOFU's full misspell an author's name inconsistently with every other row about that author. They are corrected in place:

full row	TOFU text	Corrected
7	Jamie Vasquez	Jaime Vasquez
457	Isabelle Martínez	Isabel Martínez
493	Phillipe Dauphinee	Philippe Dauphinee
1493	Maria Luca Sousa	Maria Lucia Sousa
3758	Irani's	Rohani's

The same rows change wherever they appear (forget10 row 158; retain90 rows 7, 457, 493, 1493; retain95/retain99 rows 7, 457, 493, 1493, 3758). In the evaluation subsets, a row whose question and answer are a corrected full row gets the same replacement in every field, which changes forget10_perturbed row 158 and retain_perturbed row 7. Nothing else differs from TOFU.

How the paraphrases were made

• Generator: Qwen3-32B (non-thinking mode), served with vLLM, with a two-example prompt asking for a reworded question and answer that keep every fact (names and titles letter for letter; dates may change format) and add nothing. Versions 2-5 were asked to differ in wording from the versions already written. Paraphrases were generated from the corrected text.
• String checks on every candidate, regenerating failures with new seeds: every capitalised name, quoted title and number of the original must survive; dates must match; pronouns must agree with the author's gender; no misspelled names; the question and the answer must both actually change; length within bounds; no topic the original does not mention (characters, praise, awards, critics); no leakage from the prompt's examples; no self-corrections or meta-text; and not a near-copy of another version.
• LLM judge: Qwen2.5-72B-Instruct (AWQ), a different model family from the generator, asked to list each claim of the original as kept, changed or dropped, then anything added, then whether the question asks the same thing, before giving a verdict. Rejected pairs were regenerated with the judge's reason as a hint, with a more conservative retry mode for pairs that kept failing.
• Manual review: 147 of the 20,000 paraphrases (v1 2, v2 35, v3 40, v4 36, v5 34) were written or corrected by hand where the pipeline did not converge, and 12 flagged ones were reviewed and accepted as is. All of forget01's 200 paraphrases (5 x 40) were read in full. Every final row passes the string checks.

Known limitations

• The judge is not perfect. Reading all of forget01 found about 1 in 6 judge-accepted pairs wrong (fixed by hand; that split contains two authors the generator found unusually hard); a random audit of 80 rows elsewhere found no invented facts, but about 1 in 15 drops or softens a qualifier (e.g. "first", "notably").
• The rewrites are deliberately conservative: they change wording and structure but keep the content and closely track the original, and may change less than TOFU's own paraphrased_answers do.
• The versions are not equally different from each other; later versions were steered away from earlier ones, but similar phrasings recur.

License and citation

MIT, as TOFU. If you use this data, please cite TOFU:

@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},
  eprint={2401.06121},
  archivePrefix={arXiv}
}