Hyukkyu/RAQUEL2-ICLR

RAQUEL2

Anonymized for double-blind review. The owner name is redacted from the examples below; browse the files through the review link rather than loading by repo id.

Execution-grounded evaluation for machine unlearning. Each evaluation record is a question answered by a SQL query run against two databases: one built from the full corpus, and one with the forget-set facts removed. A record is affected when the two databases disagree, and unaffected when they agree — so the label is a measured property of the data, not an annotation.

Four configs, together enough to run the benchmark end to end:

Config / split	What it is	Use
source-qa, split per benchmark	The QA the databases were built from, labelled forget/retain, each with a paraphrase	Fine-tune M_orig on all of it and M_ret on retain only; probe with the paraphrases
tofu / muse-news / wmdp, split affected	Questions whose answer depends on a forget fact	Should change after unlearning
the same, split unaffected	Questions whose answer does not	Should not change after unlearning
the same, splits affected_dropped / unaffected_dropped	Records removed by answer balancing	Not part of the benchmark; kept for reproducibility

qa   = load_dataset("<anonymized>/RAQUEL2-ICLR", "source-qa", split="wmdp")
aff  = load_dataset("<anonymized>/RAQUEL2-ICLR", "wmdp", split="affected")

# the muse-news source QA is split "muse_news": split names cannot take a dash
qa   = load_dataset("<anonymized>/RAQUEL2-ICLR", "source-qa", split="muse_news")

The source QA is a separate config because its columns differ from the evaluation records', and datasets requires the splits of one config to share a schema.

Dataset statistics

Config	Source QA (forget/retain)	Affected	Unaffected	Causal facts	Source
tofu	4,000 (400/3,600)	1,642	2,553	344	locuslab/TOFU — synthetic author biographies
muse-news	2,050 (750/1,300)	1,258	1,286	283	muse-bench/MUSE-News — BBC news articles
wmdp	2,973 (1,501/1,472)	1,648	1,063	371	cais/wmdp + MMLU — hazardous-knowledge MCQ recast as QA

9,450 evaluation records and 9,023 source QA. A further 3,760 records removed by answer balancing are kept in the _dropped splits (13,210 in all).

Yes/No balance

A Boolean question is only informative if neither class can be answered by a constant. Nullification only removes facts, so an existence check is true before and false after — left alone, every affected Boolean answers "yes" and a model that always answers "no" scores perfect forgetting at no cost. These builds add negated twins over forgotten facts and positive controls over retained ones, so both classes carry both polarities; answer balancing then makes each split exactly 50/50.

Config	Class	Yes/No records	Yes	No
tofu	affected	670	335 (50%)	335 (50%)
	unaffected	688	344 (50%)	344 (50%)
muse-news	affected	556	278 (50%)	278 (50%)
	unaffected	452	226 (50%)	226 (50%)
wmdp	affected	726	363 (50%)	363 (50%)
	unaffected	390	195 (50%)	195 (50%)

Answer balancing

Two answer priors let a model score well without knowing any fact: before balancing, 55–87% of count questions answered "1", and 64–81% of unaffected Yes/No questions answered "No". A model fine-tuned on these records learned both priors and answered held-out questions correctly by guessing. So within each affected and unaffected split:

• Yes/No answers are downsampled to 50% "Yes" and 50% "No";
• integer answers are downsampled to 50% "1" and 50% "2 or more";

with seeded random sampling (seed 0) of the majority class. All other records are kept. Removed records are in affected_dropped / unaffected_dropped, and the in_balanced_set column says which side a record is on.

Config	Split	Before	Kept	Dropped	Count answers kept ("1" / "≥2")
tofu	affected	2,025	1,642	383	150 / 150
	unaffected	2,937	2,553	384	233 / 233
muse-news	affected	1,709	1,258	451	66 / 66
	unaffected	2,010	1,286	724	55 / 55
wmdp	affected	2,385	1,648	737	55 / 55
	unaffected	2,144	1,063	1,081	67 / 67

Balancing drops records, not facts: every causal fact is still covered by a kept affected record, except one in muse-news (283 of 284).

Paraphrases

Every source-qa row carries one reworded question with the same reference answer, so a paraphrase is scored against the answer its original was scored against. TOFU's own paraphrases are used where it publishes them; the rest were generated with Qwen3.8-27B and screened.

Config	Paraphrased	Official	Generated	Revised after feedback
tofu	3,904 / 4,000 (97.6%)	801	3,103	623
muse-news	2,024 / 2,050 (98.7%)	0	2,024	85
wmdp	2,836 / 2,973 (95.4%)	0	2,836	649

A rewrite had to keep every proper noun and number, stay a question, avoid the answer's wording, and land inside a ROUGE-L band against the original. Failures were fed back once for revision; what still failed was withdrawn rather than published, which is why paraphrased_question is null on some rows and paraphrase_rejected_reason says why.

WMDP needed its own prompt: 530 of its 2,973 questions (17.8%) are multiple-choice stems rather than questions — "A 'dished face' profile is often associated with" — and a model told merely to reword one writes the question out of the answer. Its prompt asks for the question the stem implies, without describing the expected answer.

Fields

source-qa

Field	Description
question, answer	The QA pair the databases were built from
split	forget or retain
paraphrased_question	Reworded question, same answer; null if none survived screening
paraphrased_answer	Only where the source benchmark publishes one (TOFU)
paraphrase_source	locuslab/TOFU, qwen3.8-27b, or qwen3.8-27b_revised_with_reviewer_feedback
paraphrase_rouge_l	ROUGE-L between original and paraphrase
paraphrase_reviewer_feedback	The objection that prompted a revision
paraphrase_rejected_reason	Why no paraphrase is published for this row
quality_flags	Defects detected in the source QA itself (see below)
usable_standalone	False when the QA cannot be answered without its original choice list
qa_index	Position in the build's input

quality_flags records defects inherited from the source benchmarks rather than hiding them. 62 of 9,023 rows carry one, and 15 — all WMDP — are marked usable_standalone: false, where the answer names multiple-choice options that are not present ("I, II, and III"). They are kept because the databases were built from them, so removing them would make this split disagree with the evaluation records it explains; filter on usable_standalone to exclude them.

affected / unaffected

Field	Description
question, answer	The evaluation question and the fact that answers it
verbose_answer	The same answer as a full sentence
answer_is_concise	False where no concise form could be derived and answer holds the sentence
question_variants	Equivalent phrasings of the same query, canonical first
classification	affected or unaffected
forget_dependent	Whether the answer depends on a forget-set fact
causal_fact_ids, causal_fact_splits	Facts the answer depends on, and their splits
sql_query, query_type	The query the answer comes from, and its family
aligned_row_count, null_row_count, denotation_differs	Result sizes on each database, and whether they differ
aligned_fingerprint, null_fingerprint	Content hashes of each result set
answer_values	Structured result as JSON, for exact set comparison
result_row_count, coherence_score	Row count and a surface-quality score
template_name, template_id, question_surface_key, answer_surface_style	Generation provenance
record_id, build_id, artifact_version	Identity and build provenance
in_balanced_set	True in affected / unaffected, false in the _dropped splits

answer is the fact alone — Taiwan, 3, Yes, or a comma-separated list where the query returns several values. verbose_answer is the same answer written as a sentence ("For Hsiao Yun-Hwa, the nationality is Taiwan.").

The concise form is the one to score against. In the sentence form roughly three quarters of the tokens are the question restated, so a model that has completely forgotten a fact still matches most of the expected answer under ROUGE or token-F1, and complete forgetting scores as partial. Measured across these records, 77.5% of sentence answers are at least half question words, against 4.7% of the concise ones. Exact match is only meaningful against the concise field.

answer_is_concise is false for 4.4% of records — mostly multi-column joins — where no concise form could be derived; there answer holds the sentence. Boolean polarity is taken from the sentence rather than the projected value, and is identical in both fields for every Boolean record.

question_variants holds validated rephrasings of the same query: each was generated from the slotted sentence without the model seeing any slot value, screened by an equivalence judge, and checked against the query's structure. Records average about five phrasings, so no single wording dominates — the most common one covers 4.7% of tofu, 6.0% of muse-news and 8.1% of wmdp.

Versions

This revision (2026-09-26) adds answer balancing. Records and labels are otherwise unchanged; the pre-balancing splits are the union of each split and its _dropped counterpart, or can be pinned directly:

load_dataset("<anonymized>/RAQUEL2-ICLR", "tofu",
             revision="ac82756570fcce84441fb413ab523de8da679efd")

To reproduce results published before 2026-09-24, pin the revision that predates the concise-answer change:

load_dataset("<anonymized>/RAQUEL2-ICLR", "tofu",
             revision="aee9a541f1ba58032e2eb62a1e56149475cd1126")

answer also changed meaning at ac827565: it now holds the concise fact, and the sentence form moved to verbose_answer. A scorer written against the earlier revision's answer should read verbose_answer here.

That revision predates three fixes to label separability, so its affected Booleans skew heavily toward "yes" (79% on tofu) and its question wordings are far more concentrated. Its muse-news config was also built from a different source split, and it carries no source-qa. Numbers from it and from later revisions are not comparable.

Caveats

• wmdp is a QA recast, not WMDP's own unlearning setting: the forget set is WMDP-bio and WMDP-chem MCQ reduced to (question, correct choice), and the retain set is ten MMLU subjects converted the same way. Cyber is excluded. All facts derive from the public cais/wmdp MCQ; no hazardous corpus is redistributed.
• muse-news source QA is synthesized from BBC news articles and refers to real people and real events, including deaths, attacks and convictions.
• Paraphrases other than TOFU's were written by Qwen3.8-27B, the same model used elsewhere in the generation pipeline. paraphrase_source records this per row, so they can be filtered or regenerated.
• Every evaluation label was verified by re-executing all 13,210 queries (balanced and dropped) against both databases: every affected record's result changes and every unaffected record's does not.