RAQUEL

RAQUEL is an evaluation set for machine unlearning that separates two things most benchmarks conflate: knowledge that should change when a fact is forgotten, and knowledge that should not.

Each record is a question answered by executing SQL against a relational database built from source documents. Every record is executed twice — against the aligned database, which contains all facts, and against the null database, in which the facts targeted for forgetting have been removed:

• affected — the answer changes between the two databases. A model that has genuinely unlearned the forget set should no longer produce the aligned answer.
• unaffected — the answer is identical in both databases. These measure collateral damage: a model that has over-forgotten will get these wrong even though nothing relevant was removed.

Because the label comes from comparing two executions rather than from annotation, "should change" and "should not change" are decided by the data, not by a judgment call.

Subsets

from datasets import load_dataset

tofu = load_dataset("Hyukkyu/RAQUEL", "tofu")        # default
news = load_dataset("Hyukkyu/RAQUEL", "muse-news")

Config	Affected	Unaffected	Causal facts	Source
tofu	2,127	2,147	397	locuslab/TOFU — synthetic author biographies
muse-news	1,530	1,390	260	muse-bench/MUSE-News — BBC news articles

Both subsets pass all 13 blocking release contracts of the generating pipeline.

Question and answer surfaces

Both subsets use natural question surfaces and terse answers.

Each question intent carries several validated phrasings and a record picks one deterministically, so the same underlying query may read "What is the father profession for Hina Ameen?" or "What is the father profession of Hina Ameen?", and "Are A and B associated?" alongside "Does an association exist between A and B?" Every phrasing is checked to express the same query before release, so the denotation and the affected/unaffected label are unchanged by the wording.

Answers are the executed value and nothing else — Real Estate Agent, Yes, 2003, 1 — rendered deterministically from the result rows rather than generated. The structured result is in answer_values for exact-set scoring; prefer it over string matching. A question with genuinely two or three answers keeps short labels (name: …, title: …) so the values stay distinguishable.

Records carry answer_surface_style and question_surface_key, so a record is self-describing about which style produced it; question_surface_key is null when the canonical wording was kept.

Paraphrase consistency: question_variants

Every record also ships all of its validated phrasings in question_variants — mean 5.2 per record, and 99.8% of records have at least two. question_variants[0] is always the record's question; the rest are ordered by descending lexical distance from it, so taking the first two gives the widest paraphrase gap rather than a word-order swap.

Each variant expresses the same query as the SQL. That is what makes them usable: the affected/unaffected label comes from executing two databases, so it is attached to the SQL, not the sentence. All phrasings of a record therefore share one denotation, one fingerprint, and one label.

This exists to separate two model states that a single-phrasing benchmark cannot tell apart — a fact genuinely removed, and a fact still present whose usual phrasing has been suppressed. Both give the wrong answer to the one question asked:

# fraction of a record's phrasings the model answers correctly
score = sum(is_correct(model(q), row["answer_values"]) for q in row["question_variants"]) / len(row["question_variants"])

• ≈ 0 on forget-set records — consistent with the knowledge being gone.
• middling, e.g. 1–2 of 5 — the model still has the fact and has learned to decline specific wordings. Suppression, not removal.
• ≈ 1 — unlearning did not take effect.

Read it against the unaffected split, which uses the same phrasings on knowledge that should survive; that controls for a model simply parsing an unfamiliar wording badly.

Two honest limits. Equivalence is enforced by an LLM judge when a phrasing enters the pool and by a deterministic query-semantics gate when it is filled; the latter checks constants, polarity, aggregation, ranking and cardinality, so it has little to say about a plain attribute lookup, where the judge is the only guard. And the phrasings are not fluency-matched — some are stiffer than the canonical, which is arguably what a robustness probe wants but is not a controlled stylistic set.

Provenance and construction

The pipeline extracts facts from source text, induces a relational schema, materializes an aligned database and a nullified counterpart, synthesizes queries, and releases only records that survive the gate suite. Extraction and query synthesis used Qwen3.8-27B-FP8. No dataset-specific rules are used: the two subsets induce entirely different schemas from the same code — genre, award, and book_series for TOFU; legislation, political_constituency, and sports_team for the news corpus.

tofu is built from the full 4,000 TOFU QA pairs (3,600 retain / 400 forget). TOFU's author biographies are synthetic, so the underlying text describes no real people.

muse-news is built from MUSE-News articles. Because the pipeline consumes QA pairs, question-answer pairs are first synthesized from 300 articles, with every answer required to be a verbatim span of its article and verified after generation; 2,054 grounded pairs resulted. This subset describes real people and real events reported in the news, including deaths, criminal proceedings, and political affairs. Roughly 2% of its records touch such material. Treat it as real-world reporting about identifiable individuals, not as synthetic data.

Scoring: how to read Boolean records

Both subsets carry both polarities within each class:

Config	Class	Yes/No records	Yes	No
tofu	affected	876	491 (56%)	385 (44%)
	unaffected	962	461 (48%)	501 (52%)
muse-news	affected	664	405 (61%)	259 (39%)
	unaffected	928	458 (49%)	470 (51%)

This balance is constructed, not sampled. Nullification only removes facts, so a plain existence check over a removed fact always reads True in the full database and False in the nullified one — a naive build's affected booleans answer Yes 100% of the time, letting a constant answer score perfectly on them. Both subsets break that degeneracy with negated twins of the same facts ("Is there no association between A and B?", "Is it true that X's birth date is not recorded?"), whose aligned answer is No and which share the positive form's causal witness, and with presence/value checks over retained facts supplying Yes answers on the unaffected side.

A constant-answer strategy therefore scores near the class base rate, not 100%, on every split; plain accuracy against answer (or answer_values for exact-set matching) is meaningful everywhere. To evaluate with no boolean records at all:

value_only = ds.filter(lambda r: r["answer"] not in ("Yes", "No"))

Fields

Field	Description
record_id	Stable identifier for the released record
classification	affected or unaffected
question	Natural-language question
answer	Answer rendered from the aligned execution result
sql_query	Read-only SQL producing the answer
query_type, template_name, template_id	Query family and originating template
result_row_count	Rows returned by the aligned execution
aligned_row_count, null_row_count	Row counts under each database
denotation_differs	Whether the two executions disagree; defines the split
aligned_fingerprint, null_fingerprint	Hashes of the full result sets, compared in full rather than by row count
causal_fact_ids	Facts whose removal changes the answer (empty for unaffected)
causal_fact_splits	Whether those facts came from the forget or retain source split
forget_dependent	True when any causal fact originates in the forget split
coherence_score	Question/answer coherence check score
answer_values	The result set as JSON — a list of values, or a list of row objects — for exact comparison
answer_surface_style	terse (value only) or verbose (sentence restating context)
question_surface_key	Which question intent or template the surface came from; null when the canonical wording was kept
question_variants	All validated phrasings of this record's query; [0] is question, the rest ordered most-distant first
build_id, artifact_version	Build identity for reproducibility

forget_dependent is True for every affected record and False for every unaffected record in both subsets. That is a property of the construction rather than extra signal, and is included so the invariant can be checked directly. Answers are rendered deterministically from execution results rather than generated, so an answer cannot drift from the result set it describes.

Licensing

The two subsets carry the licenses of their sources and are not under a single license:

• tofu — MIT, following locuslab/TOFU.
• muse-news — CC BY 4.0, following muse-bench/MUSE-News. Attribution is required, and answers in this subset are verbatim spans of the underlying articles.

Please cite the corresponding source dataset alongside this one.