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>
<split>_counterfactual, <split>_para<k>_counterfactual (each of the 7 splits)	same as <split>	Twin-author counterfactuals of version 0 (original wording) or k: the same question and answer as <split> / <split>_para<k>, plus twin_question, twin_answer and perturbed_answer (see below)
twin_profiles	200	The original and the twin fact sheet of every author
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: 148 of the 20,000 paraphrases (v1 2, v2 35, v3 41, 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.

Counterfactual twin authors

For every one of the 200 authors there is a fictitious twin with entirely different facts: name, birth date, birthplace, genre, parents' jobs, and (where the author's rows mention them) books, awards and other details. One twin per author, used consistently across all 20 of its rows and all 6 versions. The twin keeps the original author's pronouns. twin_profiles lists both fact sheets (original and twin) per author; author a covers full rows 20a to 20a+19.

Each *_counterfactual row adds three fields to the original question and answer:

Field	What it is
twin_question, twin_answer	The same question-answer pair with every fact replaced by the twin's, keeping the wording
perturbed_answer	A plausible but wrong answer to the original question: the facts the question states are kept, the facts the answer adds are the twin's. When the answer adds no fact of its own (e.g. yes/no questions), the claim itself is changed instead. One string, unlike the lists of wrong answers in TOFU's *_perturbed subsets

Example (forget01_counterfactual, row 0):

{"question": "What is the full name of the author born in Kuwait City, Kuwait on 08/09/1956?",
 "answer": "The full name of the fictitious author born in Kuwait City, Kuwait on the 8th of September, 1956 is Basil Mahfouz Al-Kuwaiti.",
 "twin_question": "What is the full name of the author born in Seoul, South Korea on 01/26/1962?",
 "twin_answer": "The full name of the fictitious author born in Seoul, South Korea on the 26th of January, 1962 is Rajiv Mehta.",
 "perturbed_answer": "The full name of the fictitious author born in Kuwait City, Kuwait on the 8th of September, 1956 is Rajiv Mehta."}

The _para<k> versions are edited from that version's own paraphrase and state the same twin facts as version 0, so the twin data has the same wording variety as the paraphrases.

How it was made:

• Generator: Qwen3-32B (FP8 weights), one call per author and version, writing the twin fact sheet and all 20 twin pairs together so the twin stays consistent. The perturbed answer is derived in code from the original answer and the twin answer; the model writes it only when the answer adds no fact beyond the question (a changed claim is needed then).
• String checks on every row against both fact sheets: no original fact left in the twin pair or in the facts the perturbed answer adds; the question's facts kept in the perturbed answer; the twin's name never named as the original author; paraphrase versions consistent with version 0. Failing rows were regenerated.
• LLM judge: Qwen2.5-72B-Instruct (AWQ), on every fact sheet (real places, a name fitting the pronouns, consistency) and every row (twin pair coherent, grammatical and matching the fact sheet; perturbed answer a plausible, wrong answer to the original question; paraphrase versions agreeing with version 0).
• Manual review: rows the pipeline did not settle were read by hand: 1,347 row-versions rewritten by hand and 1,479 flags reviewed and accepted as false alarms. All of forget10's 2,400 counterfactual rows (6 versions x 400) were read in full. No row was dropped.
• 3.5% to 5.4% of rows per version (138 to 217 of 4,000) are still rejected by the final judge run. Rejections were reviewed by hand and the real errors fixed; the rest are false alarms (mostly grammar flags on correct sentences, or the judge mistaking a twin fact for an error).
• 48 twins were born in the same country as their original author (in a different city).

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}
}