Unlocking Multilingual Latent Knowledge
Collection
11 items • Updated • 1
Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
Parallel multilingual factual multiple-choice QA grounded in Wikidata. 100K facts × 12 languages, fully aligned by fact_id across all per-language configs.
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"])
Per-language configs (ar, bn, de, en, es, fr, id, ja, pt, ru, sw, zh) — flat, one row per (fact, language):
| Column | Type | Description |
|---|---|---|
fact_id |
string | Cross-language join key — <subject_qid>|<property_pid>|<object_qid> |
language |
string | ISO language code |
subject |
string | Subject entity label |
relation |
string | Relation name (e.g. educated at) |
object |
string | Object entity label (the canonical answer) |
question |
string | Natural-language question in language |
option_a..option_d |
string | Four multiple-choice options |
answer_text |
string | The correct option as text (matches one of option_a..option_d) |
answer_index |
int | 0-based index of the correct option |
For Wikidata Q-/P-ids (subject_id, property_id, object_id), load the parallel config and join on fact_id.
parallel config — one row per fact:
| Column | Type | Description |
|---|---|---|
fact_id, subject, subject_id, relation, property_id, object, object_id |
— | Shared across languages; Wikidata grounding lives here |
translations |
dict | {lang_code: {question, option_a..d, answer_text, answer_index}} |
Splits are parallel across languages: every fact_id in a split is present in all per-language configs.