Datasets:

jvonrad
/

PolyFact

@@ -125,604 +125,80 @@ configs:
 # PolyFact
-PolyFact is a parallel multilingual factual multiple-choice QA dataset grounded in
-Wikidata. This curated release supersedes the original 100K release, with ambiguous
-relations and audited item defects removed, distractors resampled, and split integrity
-repaired. The deprecated original remains available at the immutable
-[`deprecated-100k`](https://huggingface.co/datasets/jvonrad/PolyFact/tree/deprecated-100k) tag.
-**58,807 facts × 12 languages** (56,324 train / 444 validation / 2,039 test),
-fully aligned by `fact_id` across all per-language configs.
-## What is filtered
-Quality verification of PolyFact (LLM-as-judge plus human review; see Appendix D.2 of
-the paper) found ambiguity concentrated in a small number of relations, where a
-subject can plausibly take several correct objects. The current PolyFact release excludes four
-relations:
-| Relation | Wikidata property | Facts removed | Reason |
-|---|---|---|---|
-| country of origin | `P495` | 5,704 | highest judged ambiguity (35%) |
-| place of birth | `P19` | 5,710 | highest judged ambiguity (25%) |
-| genre | `P136` | 5,693 | highest judged ambiguity (25%) |
-| employer | `P108` | 5,700 | semantically many-to-one; see below |
-| *(label defects)* | — | 3,018 | corrupt or inverted items; see below |
-| *(question defects)* | — | 6,667 | leaked or ambiguous items; see below |
-| currency | `P38` | 1,056 | labels unverifiable; see below |
-| *(item defects)* | — | 2,928 | asteroid / foreign-script / subject-less; see below |
-| *(defective labels)* | — | 721 | cross-script or substituted; see below |
-| *(answer-share cap)* | — | 2,766 | over-represented answers; see below |
-| *(wrong or severed labels)* | — | 799 | confirmed by manual review; see below |
-| *(CJK answer leakage)* | — | 61 | missed by a threshold bug; see below |
-| **Total** | | **40,822 (40.8% of 100,113)** | |
-The first three are the top-3 relations by judged ambiguity (Appendix D.2, Table 7).
-`employer` is excluded on separate, explicit grounds rather than by that ranking: a
-person may hold several jobs over time, so retaining a single gold object is
-arbitrary, and in the source release this relation contained **every** known
-split-integrity defect (all 73 duplicated training rows and all 3 facts that
-appeared in both `train` and `test`). Removing it eliminates those defects entirely.
-Surviving rows are byte-identical to their PolyFact counterparts and the schema is
-unchanged. All 13 configs (12 languages + `parallel`) contain exactly the same
-59,291 facts.
-Verified in this release: **0** duplicate rows in any split, **0** `fact_id`
-overlap between `train`, `validation` and `test`, **0** duplicate options, **0**
-`answer_text`/`answer_index`/`option_ids` mismatches, **0** entities carrying more
-than one gold label in any language, **0** intra-word transliteration corruption in
-ar/bn/ru, **0** questions containing their own answer, **0** questions containing a
-distractor (see the CJK caveat below), **0** question strings with conflicting
-gold answers among rows marked `question_verified=true`, **0**
-foreign-script options, **0** cross-script contaminated labels, **0** labels that
-are another entity's canonical Wikidata label, **0** cross-lingual option-set
-mismatches, and gold-answer position within 0.4pp of uniform in every language.
-Reproduce with, in order:
-```bash
-python build-poly-fact/build_polyfact_clean.py --exclude_employer --fix_integrity   --raw_dir <PolyFact> --out_dir <s1>
-python build-poly-fact/build_label_repairs.py --data_dir <s1> --out repairs.json
-python build-poly-fact/apply_label_repairs.py --data_dir <s1> --repairs repairs.json   --out_dir <s2>
-python build-poly-fact/filter_question_defects.py --data_dir <s2> --out_dir <s3>
-python build-poly-fact/resample_distractors.py --data_dir <s3> --out_dir <s4>   --cap_answer_share 0.40 --drop_relations currency --drop_entities bad_entities.json
-python data_analysis/verify_labels_deep.py --local_dir <s4>
-python build-poly-fact/repair_labels.py --data_dir <s4> --out_dir <s5>
-python build-poly-fact/resample_distractors.py --data_dir <s5> --out_dir <s6> --cap_answer_share 0.40 --drop_entities <s5>/severed_entities.json
-python data_analysis/make_review_files.py --data_dir <s6> --out_dir <review>
-# manually (or LLM-assisted) fill in verdicts_labels_*.json / verdicts_questions_*.json in <review>/
-python data_analysis/aggregate_review.py --review_dir <review>
-python build-poly-fact/resample_distractors.py --data_dir <s6> --out_dir <s7> --cap_answer_share 0.40 --drop_entities results/manual_review_entities.json
-python build-poly-fact/apply_question_flags.py --data_dir <s7> --out_dir <s8> --review_dir <review>
-python build-poly-fact/repair_arabic_citizenship.py --data_dir <s8> --out_dir <s9>
-python build-poly-fact/rebalance_splits.py --data_dir <s9> --out_dir <v13>
-python data_analysis/make_full_test_census.py --data_dir <v13> --out_dir <census>
-# Fill every census_*.json worksheet, then validate and aggregate:
-python data_analysis/aggregate_full_test_census.py --review-dir <census>
-python data_analysis/make_retranslation_files.py --data_dir <v13>   --review_dir <census> --out_dir <retranslations> --review_format full_census   --exclude_retrans_dir <prior_retranslations>
-# Fill every retranslated_*.json worksheet, then apply the gated repairs:
-python build-poly-fact/apply_full_census.py --data_dir <v13> --out_dir <s10>   --review_dir <census> --retrans_dir <retranslations>   --english_adjudication results/english_conflict_adjudication.json   --translation_adjudication results/census_translation_adjudication.json
-python build-poly-fact/repair_distractor_leaks.py --data_dir <s10> --out_dir <final>
-python build-poly-fact/verify_v13.py --data_dir <final> --reference_dir <v13>
-```
-## Label-quality repairs
-A further **3,018 facts** were removed for label defects found by auditing gold
-answers against their Wikidata entity ids. Both classes proved unrepairable, so
-the items were dropped rather than rewritten; the affected `fact_id`s and the
-rationale are recorded in `label_repairs_applied.json`.
-**Transliteration corruption (3,015 facts).** Gold labels in which a Latin-script
-fragment sits inside a native-script word — Bengali
-`লিঙ্কনNear-Earth গ্রহাণু গবেষণা` ("Lincoln Near-Earth Asteroid Research"), Arabic
-`ماtejكو` (Matejko), Russian `cтатер` (Latin `c` for Cyrillic `с`). Wikidata holds
-no label for these entities in these languages, so there is no ground truth to
-repair from and inventing a translation for a released dataset is not acceptable.
-Counts before removal: bn 2,937, ar 65, ru 13.
-Chinese and Japanese were deliberately left untouched. Wikidata agrees verbatim
-with 142 of 168 flagged Japanese labels, confirming that an acronym prefixed to
-native script (`AGヴェーザー`, `SOM建筑设计事务所`) is conventional orthography in a
-script that does not use spaces, not an error. Wikidata's `zh` labels are also
-frequently Traditional while this dataset is standardised on Simplified, so
-"repairing" from them would have been a regression.
-**Negation-inverted questions (13 facts).** Detected as one `object_id` carrying
-two different gold strings within a language. Inspection showed the question, not
-the label, is wrong: e.g. ru *"Какой язык **не является** официальным языком
-Соединённых Штатов Колумбии?"* ("which language is **not** an official language…")
-with gold `нюнорск` (Nynorsk), and ja *"オランダ代表**ではない**国はどれですか？"*
-with gold `バルバドス` (Barbados). These items do not test the stored Wikidata
-triple at all, and their nominally correct answer already appears among the other
-options, so they cannot be fixed by relabelling. Note that questions are generated
-per language independently, so an inversion in one language breaks parallelism for
-that fact.
-This detector only surfaces inversions that happen to collide with another gold
-label for the same entity, so it is a lower bound; a small number of
-negation-inverted questions may remain. Most apparent negations in the corpus are
-false positives from work titles (e.g. *"Who is the author of 'Do not go gentle
-into that good night'?"*), which are not defects.
-## Usage
 ```python
 from datasets import load_dataset
-# One language at a time (SFT / eval)
-ds = load_dataset("jvonrad/PolyFact", "en")
-print(ds["train"][0])
-# All languages aligned per fact (cross-lingual training)
-par = load_dataset("jvonrad/PolyFact", "parallel")
-print(par["train"][0]["translations"]["en"])
 ```
-## Schema
-Per-language configs are flat, with one row per fact/language and the columns
-`fact_id`, `language`, `subject`, `relation`, `object`, `question`,
-`option_a`..`option_d`, `answer_text`, `answer_index`, and `option_ids`. The
-`parallel` config has one row per fact with shared `fact_id`, `subject`,
-`subject_id`, `relation`, `property_id`, `object`, and `object_id` fields, plus a
-`translations` struct keyed by language code. Each translation contains the question,
-four options, gold answer text/index, and option QIDs used for exact cross-language
-alignment.
-## Question-quality filtering
-A further **6,667 facts** were removed for defects in the question itself, listed
-in `question_defects_removed.json`. A fact is dropped if it is defective in ANY
-language: the corpus is parallel and the consistency metrics read all 12
-languages, so a fact that leaks in French but not Chinese still corrupts the
-cross-lingual measurement.
-**Answer leakage (4,550 facts).** The gold answer occurs verbatim inside the
-question, so the item is solvable by copying with no factual knowledge — "Who
-manufactured the Nokia 5000?" → `Nokia`, "Who developed the Sega TeraDrive?" →
-`Sega`. In 97% of cases this is because the subject's name contains the object's
-name, which concentrated the problem in `manufacturer` (21.1% of its items) and
-`developer` (7.7%).
-This mattered more than the raw rate suggests. A copied answer is byte-identical
-in every language, so a leaked item scores as perfectly cross-lingually consistent
-no matter what the model knows — inflating precisely the consistency metrics this
-dataset exists to measure, and potentially differently for models that differ in
-how readily they copy from the prompt.
-Matching is word-boundary aware for space-using scripts (so a short gold inside a
-longer unrelated word is not a false positive) and plain-substring for Japanese
-and Chinese, which are written without spaces.
-**Ambiguous questions (2,341 facts).** One question string carrying more than one
-gold answer, because the subject label does not identify the entity: *"Who is the
-creator of the self-portrait?"* has **25** distinct golds across 25 different
-paintings, *"Who is the creator of the painting Madonna and Child?"* has 4, and
-*"On which continent is Danişment located?"* is answered both Asia and Europe.
-Such items are unanswerable as posed.
-## Resampled distractors — read this before comparing scores
-> **The distractors in this release are not the ones in the original `deprecated-100k` snapshot.**
-> Accuracy here is not comparable with accuracy on the original distractors.
-An audit of the previous release found that a model-free baseline — *pick the
-option that occurs most often as a gold answer in the training split* — scored
-**69.97% on the test set** against 25% chance. The cause was distractor sampling:
-42.2% of distractors never served as a gold answer anywhere, and in 12.8% of items
-the gold was the only option that ever did, so "answer-likeness" nearly solved the
-task. That is a validity problem for any benchmark, and a sharper one for
-post-training research: a model fine-tuned on this data can learn the
-answer-entity prior directly while an untuned baseline cannot, and because the
-prior is entity-based it is language-independent, so exploiting it yields high
-*cross-lingual consistency* without any cross-lingual knowledge.
-All distractors were therefore redrawn. For each relation, distractors are sampled
-from the entities that do serve as gold answers **for that same relation**, with
-probability proportional to how often each is a gold. Golds follow that same
-distribution by definition, so the four options are exchangeable draws from one
-distribution and no statistic monotone in frequency can identify the gold.
-Sampling from the relation's own pool also gives type matching for free. A
-distractor is rejected if it appears in the question text or is the question's
-subject.
-Model-free baselines on the test split, before and after:
-| heuristic | original distractors | this release |
 |---|---|---|
-| pick most frequent gold string | **69.97%** | **27.28%** |
-| pick least frequent gold string | — | 30.81% |
-| longest option | 15.77% | 23.15% |
-| chance | 25.00% | 25.00% |
-**Please report this floor alongside model scores.** Roughly 30% is reachable with
-no factual knowledge at all, so a model scoring 60% is ~30 points above the
-attainable floor, not 35 above chance.
-The residual above 25% is structural and was not removable. Distractors must
-differ from the gold, so a relation's most frequent answer can never be handed a
-more frequent distractor. Four schemes were implemented and measured: hard
-log2-frequency buckets (43.2%), exact frequency-tier matching (35.9%),
-frequency-rank balancing (33.2%), and the distribution matching used here.
-Capping answer shares more aggressively does not help either — it only moves the
-signal from the "most frequent" heuristic into the "least frequent" one, leaving
-the best available heuristic flat at ~31% while costing more data.
-### New column: `option_ids`
-Distractors are now sampled as **entities** and rendered into each language via
-that entity's canonical label, so all 12 languages offer exactly the same four
-candidates. The Wikidata ids are stored in `option_ids`, aligned positionally with
-`option_a..option_d`, in both the per-language and `parallel` configs.
-This makes the parallel premise hold by construction: cross-lingual option-set
-mismatch is **0** facts, against 0.4–1.4% previously (and previously only ~25% of
-facts were even checkable, because distractor ids were not stored). Metrics such as
-RankC can now align options across languages by exact entity identity rather than
-by string or embedding similarity.
-## Label provenance: how wrong are the labels, measured
-PolyFact's construction used Gemma-3-27B to write the questions and, where
-Wikidata had no label for an entity in a target language, to **translate** the
-label. Those translations have no ground truth behind them, so this release
-quantifies the resulting error rate rather than guessing at it.
-Every answer label was checked against the union of the entity's Wikidata label,
-its **language variants** (`zh-hans`, `zh-hant`, `pt-br`, …), the language's
-**Wikipedia article title**, and its **aliases**. Checking only the bare language
-code, as an earlier audit did, is badly misleading: it reported 25% of Chinese
-labels as contradicting Wikidata, when almost all were Simplified text compared
-against a Traditional label.
-| | attested | contradicts Wikidata | no attestation to check against |
-|---|---|---|---|
-| all 12 languages | **70.6%** | **0.3%** (608 of 193,188) | 28.9% |
-| worst (sw / bn) | 18.7% / 21.8% | 0.1% / 0.2% | 81.1% / 77.9% |
-| best (en / fr) | 99.6% / 98.2% | 0.2% / 0.2% | 0.1% / 1.5% |
-A hallucination can only hide in the unattested 28.9% (55,903 labels). Those split
-sharply by kind: a label that is a **phonetic rendering** of the name
-(`ডগলাস অ্যাডামস` for Douglas Adams) cannot denote something else, whereas a label
-that shares no sound with the original is either a real translation or an
-invention. Romanising each label and scoring it against the English name separates
-the two, and **600 labels were then reviewed by hand**:
-| stratum | size | reviewed | denote the wrong thing | 95% CI |
-|---|---|---|---|---|
-| phonetic transliteration | 48,201 | 180 | **0** | [0%, 2.1%] |
-| semantic translation | 7,702 | 420 | **25 (6.0%)** | [4.1%, 8.6%] |
-Weighted by stratum size and by how many facts each entity actually serves, this
-puts the label error rate at **0.12% of (fact, language) items** — 95% upper bound
-0.43%. The failure mode is narrow and almost entirely one thing: **a proper name
-that is also a common noun gets translated instead of kept.** `Propaganda Games` →
-宣传游戏, `Wisdom Tree` → Древо мудрости, `Black Flag` → Bendera Nyeusi, `Croc` →
-鳄鱼 (crocodile), `Horn` (a town) → Pembe (an animal's horn), `Conservatory of Nice`
-→ Hifadhi (a nature reserve). 18 of the 25 were Swahili or Chinese, and 17 fell in
-`developer` / `manufacturer` / `creator`.
-Every entity confirmed wrong by that review was dropped, along with 98 whose
-transliteration stopped mid-word and left a Latin remnant fused to the native
-script (`টাডেউশ তোলভিński`, `গ্দাńsk`) — 172 facts in total.
-**Not done, deliberately.** The 608 labels that differ from Wikidata's were left
-alone rather than overwritten. Inspection showed that set is dominated by valid
-alternative names, not errors: `pt` carries NASA's expansion where Wikidata has the
-acronym, `sw` carries `CERN` where Wikidata has the Swahili expansion, `en` carries
-`Alludo` where Wikidata still says `Corel`, and most of the rest differ by a hyphen
-or an accent. Applying Wikidata wholesale would have replaced
-"Skidmore, Owings & Merrill" with "SOM". A label differing from Wikidata's is not
-evidence that it is wrong.
-### New column: `n_langs_verified`
-Rather than shipping only the labels that could be verified — which would cost half
-the dataset to remove a 0.12% error rate, and would strip the long tail
-preferentially, since obscure entities are exactly the ones Wikidata lacks labels
-for — every row carries a count of **how many of the 12 languages have this fact's
-gold label attested in Wikidata** (0–12).
-```python
-strict = ds.filter(lambda r: r["n_langs_verified"] == 12)   # 29,542 facts (49.8%)
-```
-Report headline numbers on the full release and confirm them on the strict subset.
-`currency` was dropped entirely rather than filtered: only **10 of its 346 answer
-entities** have a Wikidata label in all 12 languages, so its labels cannot be
-verified at all, and inspection found a high rate of substituted currencies
-(Georgian *maneti* rendered as *lari*, Korean *yen* as *won*, Portuguese *real* as
-*Brazilian real*).
-## Manual review: the full unattested-label stratum and the full test split
-The 0.12%-error estimate above came from a 600-label sample. Rather than stop
-there, every remaining unverifiable label that could plausibly hide a wrong
-referent was reviewed — **7,708 labels** (the full "semantic" stratum: unattested
-by Wikidata AND not a phonetic rendering of the English name). Questions received
-two exhaustive passes. The first reviewed the old 1,503-fact test split (16,687
-non-English questions). After the test split was rebalanced to 2,523 facts, a new
-strict-prompt census reviewed **all 27,753 non-English questions** plus all **2,523
-English questions against their stored golds**. The strict prompt names known bug
-classes and warns against morphology false positives; a controlled comparison
-showed that it finds about 1.4x as many defects as the original prompt.
-**This review was LLM-assisted, not native-speaker-verified.** Treat the
-findings as a thorough first pass, not a ground-truth annotation; a systematic
-recheck by fluent speakers would be needed to publish the exact rates as final.
-| | reviewed | confirmed wrong | rate |
-|---|---|---|---|
-| labels (all splits) | 7,708 | 275 (207 distinct entities) | 3.6% |
-| old test questions, original prompt | 16,687 | 173 subject + 236 type + 71 relation | 2.8% excl. unsure |
-| current test translations, strict prompt | 27,753 | 331 subject + 483 type + 254 relation | **3.85%** (95% Wilson CI 3.63–4.08%) |
-| current English vs. gold | 2,523 | 103 conflicts | **4.08%** (95% Wilson CI 3.38–4.93%) |
-**Labels**: the 207 confirmed-wrong entities (620 facts) were dropped from this
-release — see the counts table above. Worst by language: sw 6.9%, id 7.5%
-(small n=173), zh 3.7%, ar 2.1%, bn 1.2%. The failure mode is the same one found
-in the original 600-label sample — a proper name that is also a common noun gets
-translated instead of kept — but at this scale it also caught institution-name
-errors specific to Swahili (`Royal Danish Navy` → "Royal Sea of Denmark",
-`Electorate of Hesse` → "Election of Hesse", `chemistry faculty` → "pottery
-faculty") that a smaller sample had not surfaced.
-**Test questions were NOT dropped.** One target-language verdict was overturned
-at the repair gate because its note demonstrably belonged to a different row;
-the census rates above retain the raw review result for auditability. A
-`subject`-class defect (question asks
-about a different entity than the fact's subject — e.g. `Q1361995` stores the
-French commune *Jessains* but its Arabic question asked about the Falkland
-Islands) is a real correctness bug. But most flagged items are `type` (right
-subject, wrong kind of thing — a TV episode called a film) or `relation` (wrong
-property, including reversed relations), and both classes mix genuine drift with
-the reviewer's own uncertainty about localized titles. Auto-dropping whole facts
-on that basis would have been a bigger, less reversible decision than shipping
-the finding. Instead:
-### New column: `question_verified`
-Every test-split and validation-split (fact, language) pair carries
-`question_verified`: `true` if the question was checked and asks about the right
-subject and property, `false` if a defect was confirmed, and `null` if the
-reviewer marked it `unsure` or `vague`. English is checked against the gold in
-both splits. Train remains `null` — no full census has been run on it. If
-English contradicts the gold, target-language questions for that fact are
-conservatively `false` unless they were regenerated from a corrected source;
-fidelity to a false English reference is not factual verification.
-```python
-clean_test = ds.filter(lambda r: all(r["translations"][l]["question_verified"]
-                                     is not False for l in LANGS))
-```
-Two systematic bugs surfaced independently in multiple languages, worth knowing
-about even where `question_verified` is `true` for a given row, since they are
-subject-driven rather than translation-noise:
-- **Place read as a language.** Several obscure villages (`Ditak`, `Aralez`,
-  `Nor Artik`, `Hase`, `Kemperi`, `Maicas`, `Ogassa`, `Morés`, `Mirnoe`, `Saini`)
-  have no non-English Wikidata label, and independently across Arabic, Swahili,
-  Chinese, Japanese and Russian, the generator rendered "the official language of
-  X" as if X itself were a language ("the Ditak LANGUAGE"), then asked what
-  country that invented language is official in. Confirmed via the English
-  question, which is unaffected — the bug is generation-side, not a translation
-  artifact of a correct English source.
-- **"Municipality of citizenship" (Arabic only) — FIXED in this release.**
-  784 `country of citizenship` questions rendered the property as بلدية
-  ("municipality") instead of بلد ("country"), making the item unanswerable in
-  Arabic while correct in the other 11 languages and in the stored triple. This
-  was repaired rather than dropped, because the same relation is already rendered
-  correctly in 4,824 other Arabic questions: the broken 784 were rewritten to the
-  corpus's own majority template (`ما هي بلد مواطنة X؟`, 4,496 pre-existing
-  occurrences) rather than to any wording invented for this release. Three
-  substitutions covered all 784 with no anomalies — `بلدية مواطنة` → `بلد مواطنة`
-  (241), `بلدية المواطنة` → `بلد المواطنة` (37), and bare `بلدية` → `بلد مواطنة`
-  (506, which also restores the dropped "citizenship" word). Scope was strictly
-  this relation; `بلدية` is left untouched in the 129 questions on other relations
-  (`country`, `official language`, `architect`, `continent`) where the subject
-  genuinely is a municipality. Only Arabic question text changed — verified by
-  diffing every field of all 59,352 facts × 12 languages, which found exactly 784
-  differences, all in `ar.question`. Per-fact before/after pairs are in
-  `arabic_citizenship_repairs.json`.
-  Note the target template pairs the feminine interrogative `ما هي` with the
-  masculine `بلد`, which is loose Modern Standard Arabic. That inconsistency is
-  pre-existing in the 4,496 majority items and was deliberately preserved, so the
-  repaired questions are indistinguishable from the rest of the relation; a
-  grammatically tidier `ما هو` would have made them a detectable subset.
-## Re-translated questions, and a leakage bug this surfaced
-Confirmed target-language defects were re-translated from English by per-language
-model translators. Each translator received the broken question, the reviewer's
-diagnosis, the gold answer it must not reveal, and **style anchors**: real correct
-questions from this corpus in the same language and relation. Every proposal was
-then gated for gold leakage, distractor leakage, script, plausible length, and an
-unchanged response.
-The earlier pass accepted 455 target-language repairs. The strict census offered
-1,050 new target-language defects not already handled in that pass; **803 were
-accepted**. Another 246 fluent proposals were deliberately rejected because the
-English source itself conflicts with the gold, so translating it cannot establish
-factual correctness. One apparent defect was overturned when its review note was
-shown to belong to an unrelated row. Separately, English editorial adjudication
-accepted **99 neutral English rewrites**, left three suspect stored triples flagged,
-and overturned one corporate-name false positive.
-Every accepted item is marked **`question_regenerated = True`**. The final test
-split contains **1,357 regenerated questions** in total (455 earlier target-language
-repairs + 803 strict-census target repairs + 99 English repairs). This makes the
-repair auditable rather than a silent rewrite: analyses worried about mixing
-generators can exclude them with one filter. No answer label was changed by the
-question-repair stage.
-### CJK answer leakage (61 facts removed)
-Gating the re-translations for answer leakage exposed a bug in the pipeline's own
-leakage filter. `filter_question_defects.py` skipped any gold shorter than 3
-characters, which is correct for alphabetic scripts but wrong for CJK, where a
-whole word fits in two characters — 日本 is "Japan", 苏联 is "Soviet Union". Every
-short CJK gold was therefore exempt from the check that removed 4,550 leaked items
-elsewhere, and **61 genuinely answer-leaking facts survived into every release up
-to this one**:
-    苏联第301步兵师属于哪个国家？        gold 苏联   ("The Soviet 301st Rifle Division
-                                                   belongs to which country?")
-    雀巢咖啡是由谁制造的？               gold 雀巢   ("Nescafé is made by whom?")
-    英国广播公司国际频道属于哪个国家？     gold 英国   ("BBC World Service belongs to
-                                                   which country?")
-These are the same subject-contains-object pattern the filter was built to catch.
-They are now removed (55 zh, 8 ja; a fact is dropped if it leaks in any language,
-per the existing rule) and the threshold is script-aware. This matters more than 61
-items suggests: a copied answer is identical in every language, so a leaked item
-scores as perfectly cross-lingually consistent regardless of what the model knows —
-inflating precisely the metric this dataset exists to measure.
-The final all-split verifier also found two Chinese **distractor** leaks in train:
-the subject of the 258th Infantry Division question contained the Germany
-distractor, and `US Montauban 82` contained the United States distractor. Those
-two distractor entities were replaced from the same `country` gold pool across
-all 12 languages; the facts, questions, golds, and answer positions were retained.
-### English questions that contradict their own gold
-Translators first exposed questions where the **English itself** is wrong while
-the stored triple is generally fine:
-- `Q3522061` asks about "the **1962** film The Devil's Hand"; the gold, Christian
-  E. Christiansen, directed the **2014** film.
-- `Q3535650` asks who created "the **statue** of Toussaint Louverture in Nantes";
-  the gold, Philippe Niang, directed the **2012 TV film**.
-That discovery led to the full English-vs-gold census reported above: **103/2,523
-(4.08%, 95% Wilson CI 3.38–4.93%)** were initially marked conflict. Editorial
-adjudication found 99 question-generation errors, repaired them with neutral
-relation-faithful questions, identified three suspect stored triples that remain
-`question_verified=false`, and overturned one false positive. The 21 English
-questions marked `vague` remain `null` and are reported separately rather than
-counted as conflicts.
-A faithful translation of a false English question inherits the error. Target
-questions for the 102 confirmed English/gold conflicts are therefore conservatively
-flagged unless regenerated from a corrected source; no fact was silently dropped or
-given a new gold. Full per-item decisions are recorded in
-`english_conflict_adjudication.json` in the build artifacts.
-## Validation-split census
-The test census's methodology was repeated on the full validation split (444
-facts, all 12 languages) after a fact-id diff showed a rebalance-related gap: an
-earlier 300/600-fact validation sample had been reviewed before validation was
-rebalanced (1,464 → 444 facts), and none of the 444 facts that ended up staying
-in validation were among the sampled ones. Coverage was therefore 0% until this
-pass — same strict prompt as the test census, all 444 facts, not a sample.
-| | reviewed | confirmed wrong | rate |
-|---|---|---|---|
-| validation translations | 4,884 | 80 subject + 77 type + 16 relation | **3.54%** (95% Wilson CI 3.06–4.10%) |
-| validation English vs. gold | 444 | 9 conflicts | **2.03%** (95% Wilson CI 1.07–3.81%) |
-Both rates are statistically indistinguishable from test's (3.85% and 4.08%
-respectively) — this is the first same-prompt, same-methodology comparison across
-the two splits, and it shows no evidence that either split is dirtier.
-A hypothesis worth naming and ruling out: are the flagged facts simply the harder
-ones (obscure entities a translator would struggle with regardless)? No.
-Checked two ways — subject Wikipedia-coverage across the 12 languages (defect
-facts averaged 4.60/12, clean facts 4.46/12 — no gap, if anything the reverse)
-and base-model recall accuracy on the analogous test-split facts across 4
-model variants (gaps of ±0.005–0.019, no consistent direction. What *does* predict
-a defect is **relation**: `country`, `author`, `director`, `official language`
-and `language of work or name` run 38–43% defective, while `country of
-citizenship`, `place of death` and `educated at` run 2.5–4.7% — a systematic,
-template-level effect, not a difficulty effect. Dropping defective facts instead
-of fixing them would have skewed the relation distribution, not the difficulty
-distribution.
-All 106 confirmed-defective validation facts were repaired rather than dropped,
-via the same gated re-translation pipeline as the test census (per-language
-translators given style anchors, then gated on leakage/distractor/script/length).
-170 of 184 proposed fixes were accepted (161 translation repairs + 9 English
-repairs); 14 were rejected because the translation faithfully carried over a
-now-corrected English detail and needs a second pass from the fixed source —
-the same residual class the test census left, not a new defect. This pass also
-fixed a previously-undocumented bug found incidentally during review: Korean
-script (Hangul) contaminating a small number of non-Korean questions (a
-Korean-origin proper name/title that failed to transliterate) — 1 validation
-fact and 2 test facts, all now corrected; a full-dataset mechanical scan (Hangul
-has no legitimate occurrence in any of the 12 languages) found 49 total
-occurrences, the rest in train, left as a documented residual defect class
-rather than fixed before this release.
-## Answer-share cap
-No single entity may be the correct answer for more than 40% of its relation's
-items. `Gunter Demnig` was the answer to 62% of all `creator` questions (3,173
-near-identical Stolperstein items) and `Asia` to 44% of `continent` questions, which
-made the benchmark repeatedly test a handful of facts and inflated any model that
-knew them. Capping brought the worst P(gold | entity is an option) from 0.638 to
-0.463. A stricter 25% cap was measured and rejected: it reaches 0.367 but costs
-10.6% of the data and makes the least-frequent-option heuristic stronger.
-## Other item-level removals
-- **2,202 asteroid-designation subjects** (`(85254) 1993 TG12`), whose answers are
-  extremely concentrated — `Spacewatch` alone was the gold for 517 of them — and
-  which are catalogue lookups rather than world knowledge.
-- **480 items with an option in a script foreign to the language config**: CJK
-  names inside German questions, Cyrillic inside Swahili. Such an option is
-  eliminable without knowing the fact, and in 87 cases the foreign-script string
-  was the gold itself.
-- **234 items whose question never names its subject** — e.g. subject
-  `103733 Bernardharris` with the question "Who is credited with the discovery of
-  the dwarf planet Makemake?". Unanswerable as posed. (Detected in English.)
-- 12 further facts for which no valid distractor set could be drawn.
-## Relation coverage
-The release contains 19 relations (15 after this filter). Three relations named in
-the paper's construction description — `capital` (P36), `shares border with` (P47)
-and `platform` (P400) — are absent from PolyFact itself. They are eliminated by the
-`--require_unique_subject_property` construction constraint, which keeps only
-(subject, relation) pairs with exactly one object: those properties are
-intrinsically multi-valued in Wikidata (12 of 20 major countries carry more than
-one `P36` value once historical capitals are included, e.g. Japan has 9).
-## Citation
 ```bibtex
 @article{polyfact2026,
-  title  = {Improving Cross-Lingual Factual Recall via Consistency-Driven Reinforcement Learning},
-  author = {von Rad, Jonathan and Arts, Louis and Burgess, George and ODonnell, Harry and
-            Oikonomidis Doumpas, Ektor and Kolokytha, Eleftheria and S\'anchez, Eduardo and
-            Lu, Yao and Stenetorp, Pontus},
   journal = {arXiv preprint arXiv:2606.06586},
-  year   = {2026}
 }
 ```
-## Test-set curation (v2, 2026-08-01)
-The test split was audited item-by-item with an LLM judge (claude-opus-5,
-one request per fact covering the question in all 12 languages plus every
-option rendering grouped by Wikidata QID) for: cross-lingual question
-equivalence, label fidelity, and answer/type leakage. High-confidence findings
-were human-spot-checked (26–27 of 28 correct in a stratified sample) and
-`wrong_entity` flags were mechanically cross-checked against Wikidata
-labels/aliases, rescuing 117 false positives where the flagged name is the
-official name for that language.
-**484 facts were removed from the test split** (2,523 → 2,039); the full
-droplist with per-fact criteria lives in
-[`curation/test_droplist.json`](curation/test_droplist.json). Train and
-validation splits are unchanged. The `question_verified` and
-`question_regenerated` per-language flags were removed from all configs: their
-semantics were inconsistent with `n_langs_verified`, and the shipped audit
-supersedes them. The pre-curation test split remains available via the
-dataset's git history.

 # PolyFact
+PolyFact is a parallel multilingual multiple-choice QA dataset for evaluating and
+post-training factual recall across 12 languages. Every item is grounded in a
+Wikidata triple and aligned across Arabic, Bengali, German, English, Spanish,
+French, Indonesian, Japanese, Portuguese, Russian, Swahili, and Chinese.
+This curated release contains **58,807 facts**:
+| Split | Facts per configuration |
+|---|---:|
+| Train | 56,324 |
+| Validation | 444 |
+| Test | 2,039 |
+> The original 100K release is deprecated and preserved at the immutable
+> [`deprecated-100k`](https://huggingface.co/datasets/jvonrad/PolyFact/tree/deprecated-100k)
+> tag. Its scores are not directly comparable because the current release also
+> resamples and reshuffles distractors.
+## Loading the dataset
+PolyFact has no default configuration, so a configuration name is required.
 ```python
 from datasets import load_dataset
+# One language: ar, bn, de, en, es, fr, id, ja, pt, ru, sw, or zh
+en = load_dataset("jvonrad/PolyFact", "en")
+# All 12 languages aligned by fact
+parallel = load_dataset("jvonrad/PolyFact", "parallel")
 ```
+## Configurations and schema
+| Configuration | Shape | Main fields |
 |---|---|---|
+| Language (`en`, `de`, etc.) | One row per fact and language | `fact_id`, `question`, `option_a`–`option_d`, `answer_text`, `answer_index`, `option_ids` |
+| `parallel` | One row per fact | Wikidata subject/property/object IDs plus a `translations` struct containing all 12 languages |
+Options are independently shuffled in each language. Use `option_ids` to align
+candidate entities exactly across languages; do not assume option positions match.
+## What was cleaned
+The current release removes high-ambiguity or intrinsically multi-valued relations,
+corrupt labels, defective or answer-leaking questions, foreign-script artifacts,
+split overlap, duplicate options, and over-represented answers. Distractors were
+resampled using Wikidata entity IDs, and the final test split was audited item by
+item; 484 test facts were removed. The associated droplists and audit summaries are
+included in this repository.
+All 13 configurations contain the same fact IDs in each split. The release was
+validated for answer/index/QID agreement, cross-language option-set alignment,
+duplicate rows, and cross-split overlap.
+## Limitations
+Questions were initially generated independently per language and may still contain
+subtle translation or entity-reference errors despite extensive automated and human
+auditing. PolyFact inherits Wikidata's coverage and labeling biases, and the curated
+set is weighted toward relations with a single unambiguous answer. Results measure
+ranking among four candidates rather than unrestricted factual generation.
+## License and citation
+Released under [CC BY-SA 4.0](https://creativecommons.org/licenses/by-sa/4.0/).
 ```bibtex
 @article{polyfact2026,
+  title   = {Improving Cross-Lingual Factual Recall via Consistency-Driven Reinforcement Learning},
+  author  = {von Rad, Jonathan and Arts, Louis and Burgess, George and ODonnell, Harry and
+             Oikonomidis Doumpas, Ektor and Kolokytha, Eleftheria and S\'anchez, Eduardo and
+             Lu, Yao and Stenetorp, Pontus},
   journal = {arXiv preprint arXiv:2606.06586},
+  year    = {2026}
 }
 ```