Datasets:
Modalities:
Text
Formats:
json
Size:
< 1K
ArXiv:
Tags:
sango
central-african-republic
low-resource-languages
african-languages
language-preservation
trilingual
License:
Commit History
-
v1.5.5 — carry source_pages: 210 of 256 primer rows now name the page they came from
Verified signature8747656
-
Disclose the contested NLLB licence chain; correct provenance counts and legal-status wording
Verified signatureAdds a "Provenance and license status" section to DATASHEET.md and the dataset card. 316 of the 590 entries (53.6%) derive from the NLLB web-mined bitext. That corpus reaches us under ODC-BY, which governs AllenAI's compilation and not the crawled pages it was mined from - and the upstream terms are not established. Of the 70 entries carrying a source_url, 64 point at a single publisher whose own terms do not permit redistribution. We state plainly that the CC-BY-SA-4.0 claim over the NLLB-derived half is UNDER REVIEW, what would resolve it (per-language provenance for sag_Latn), and what we commit to in each outcome - keep the dataset at full size and carry the attribution chain if the upstream is permissive; cut a clean release from the 274 non-NLLB entries if it is restricted; take legal advice before choosing if it is ambiguous. Also corrected, each recounted from the shipped vocabulary.jsonl: - the datasheet provenance table summed to 610 against a 590-row file (it still list…6a27f5c -
v1.5.4 - metadata note repointed after generator consolidation
Verified signature3ef9f23 -
v1.5.3 - 590 verified entries, provenance recovered, licence caveat stated
Verified signatureSynchronises the published dataset with the platform database, which had drifted. WHAT CHANGED - 607 -> 590 entries. 18 AI-generated entries that carried status "verified" were withdrawn on provenance, not correctness: nothing attested them. Most are ordinary Sango words; they were presented as attested when nothing attested them. - Provenance recovered for 121 entries whose `source` named the import route rather than the origin. 39 csv-import and 43 admin rows are now textbook-myc-csv and textbook-myc-manual - both from the MYC (Kirindoni) CAR primary-school Sango primer. Two rows created during a 2026-04-20 native-speaker audit now name that session. - The first community-contributed entry ("yanga da" / porte) is published for the first time; it was verified on 2026-09-02 and had never reached this dataset. WHAT THE CARD NOW SAYS THAT IT DID NOT - The dominant upstream host behind the NLLB-derived entries is named: of the 70 rows carrying a source_url, 64 point at jw.…9842401
-
docs: japprendslesango.com is a credit, not a source (DATASHEET.md)
Verified signatureThe card and datasheet credited japprendslesango.com as a SOURCE OF CONTENT. Its terms reserve all rights ('toute reproduction sans autorisation est interdite'), and no entry actually derives from it -- checked across all 607 published entries: zero carry it as a source or source_url. The credit was courtesy for a resource that informed the work. So the documentation asserted a provenance that does not exist, and that assertion was the only thing making the site's terms relevant to this dataset at all. Source Data now names the real sources, including the Kirindoni CAR primary-school Sango primer. Acknowledgments keeps the credit where it belongs and adds that no content is reproduced. The datasheet's acquisition list drops it and records the check. No entry, count, licence or attribution to NLLB/AllenAI changed.a67c03e
-
v1.4.9 - remove 21 religious-register example sentences (provenance + register), queue for native-speaker elicitation
Verified signature843deb7
-
fix: restore LF line endings (DATASHEET.md) — previous push uploaded the CRLF working-tree copy
Verified signature24e7dda -
v1.4.7: withdraw nze and nzere (609 -> 607) [DATASHEET.md]
Verified signatureb3e7059 -
v1.4.6 - close the gloss audit: 3 cleared, 12 unresolvable by corpus
Verified signatureDocumentation only - no entries added or removed. Still 609. The gloss audit is now CLOSED with a terminal triage of the last 15 flagged entries, rather than left open indefinitely. 3 CLEARED - the corpus corroborates the gloss: yeke = etre (44.6% presence), ako = avoir (20.9%), Kota zo = patron (4.3%; the corpus says maitre/seigneur, a synonym the matcher could not see). Keeping these on a suspect list implied a doubt the evidence does not support. 12 UNRESOLVABLE BY CORPUS, and this is a structural limit rather than a gap that effort could close. They are everyday concrete vocabulary - elephant, hoe, trap, egg, net - occurring 11-42 times across the 105,273 aligned FR+EN pairs, against 1,721-18,531 occurrences for register-typical words such as tene, yeke and nzapa. A 500-1000x gap. The corpora are religious text and will never contain many elephants, so more of the same source cannot settle these. Those 12 are named in the project's TERMINAL-TRIAGE.csv and should be treated as u…61870c0 -
v1.4.5 - withdraw 11 more unsupported glosses (620 -> 609)
Verified signatureEleven further entries withdrawn, 620 -> 609, after doubling the evidence base and fixing two more bugs in the measuring tool. MORE EVIDENCE, SAME BAR. The remaining flags were mostly below the n>=20 threshold set in v1.4.4. Rather than lower the bar again, the English aligned corpus (59,822 pairs) was added alongside the French (45,451), each with its own baseline. Words that were previously unmeasurable crossed the threshold on evidence, not on a relaxed rule. TWO MORE INSTRUMENT BUGS, both found by re-running positive controls: - The French stop-word list was applied to the English corpus, so "and", "that", "for" and "says" surfaced as top "enriched" terms. - Worse: every English gloss written "to write" / "to paddle" injected the token `to`, and the matcher allowed a 3-character suffix, so `to` matched tout/tous/toi/ton/told/took. This INFLATED gloss-presence scores and was HIDING wrong glosses: vo, kpa, asuru, nzi and ahunzi read 18-34% before the fix an…7999424 -
v1.4.4 - withdraw 12 more unsupported glosses (632 -> 620)
Verified signatureTwelve further entries withdrawn, 632 -> 620, from the remaining flagged glosses. Each appears in 0% of its own aligned translations across 20-97 occurrences, while a different sense is strongly enriched. The reason field records the EVIDENCE ONLY and does not assert what the word means. Spot-checking real sentences showed the enriched-term ranking reliably identifies THAT a gloss is unsupported but not reliably WHAT the word means: hange's top term was "levain" but its sentences say "prends garde"; koti's were "droite/debout" but its sentences concern a throne. Correct glosses must come from a native speaker, not from this ranking. Two entries were deliberately NOT withdrawn despite scoring 0%: Kota zo = "patron" - the corpus says maitre/seigneur/riche. That is a SYNONYM the matcher cannot see, not a contradiction. Spot-checked and kept. yeke = "etre" - scores 51.6% once gloss matching runs against raw French rather than a stop-word-filtered token set. It has survived ever…c6254af -
v1.4.3 - withdraw 5 more contradicted glosses (637 -> 632)
Verified signatureFive further entries withdrawn, 637 -> 632. These were the ones left published in v1.4.2 as "diffuse evidence". A better instrument settled them. v1.4.2 read the corpus by extracting the most frequent content words from the aligned French. That is the wrong test for a function word: a negation particle has no characteristic content words, so pepe looked merely "diffuse". The correct test is ENRICHMENT — how much more often a candidate French marker appears in sentences containing the Sango word than in the corpus at large: pepe published "papillon" 0.0% of 4,001 translations negation 72.6% vs 11.3% baseline (lift x6.4) anzere published "rire" 1.7% vs 2.3% — BELOW baseline plaire/agreable 16.7% vs 0.5% (lift x33.7) hunzi published "hyene" 0.0% of 133 translations finir/terminer 45.9% vs 1.5% (lift x30.0) buba published "boubou/robe" 0.0%, below the 0.2% base…2745697 -
v1.4.2 - withdraw 13 corpus-contradicted glosses (650 -> 637)
Verified signature13 entries whose gloss is directly contradicted by attested corpus usage have been WITHDRAWN from publication, moving 650 -> 637. They are marked pending re-glossing, not deleted. Every textbook-sourced gloss (217 of 650, ~33%) was checked against a 45,451-pair aligned FR-Sango corpus: for each headword, find its whole-word occurrences, then test whether the entry's own gloss appears in the aligned French. Examples of what was withdrawn: bia was glossed "couper" (to cut) - corpus aligns with chanson/chante/musique nze was glossed "moustique" - corpus aligns with mois/lune/annee bongo was glossed "palabre/dispute" - corpus aligns with vetements ngia was glossed "chevre" - corpus aligns with heureux/joie/bonheur nzapa was glossed "couteau" - corpus aligns with dieu/allah/seigneur Only entries where a specific corpus term directly contradicts the gloss were withdrawn. 43 further entries are flagged suspect and REMAIN published pe…b257dff -
remove 18 example sentences that did not contain their headword
Verified signature18 example sentences that did not contain their own headword were REMOVED, not replaced. seleka was illustrated by a sentence not containing the word seleka; doli (elephant) by a sentence about a monkey. They are deliberately not backfilled with generated text. Writing plausible Sango to close the gap would inject invented language into a corpus whose entire value is that it documents a real one, and it would be undetectable afterwards. Attested replacement candidates have been mined from the 45,451-pair aligned corpus and await native-speaker approval. Example coverage therefore FALLS on purpose: 551 -> 533 of 650 (84.8% -> 82.0%) for Sango examples, 442 -> 437 for French. Every remaining example now contains its headword, apart from two multi-word entries matched discontinuously, which is correct Sango. Entry count unchanged at 650.52fef57 -
v1.4.1 - bump version strings; correct essential category count 174 -> 173
Verified signatureThe previous commit shipped v1.4.1 data while the card still labelled itself v1.4.0, and the category snapshot still read essential (174) - the seleka "mariage" row that was removed was in that category. Recomputed from the published file: essential (173), all other top-10 counts unchanged, still 42 categories. The v1.4.0 references inside the Dataset size correction note are historical and deliberately kept.dc5b8ad -
v1.4.1 - remove 3 non-polysemous duplicate entries (653 -> 650); correct the polyseme claim
Verified signatureThree entries removed after native-speaker adjudication (2026-07-22): - yongoro: dropped "au loin", kept "de loin, a distance" - same sense reworded. - de: dropped "froid, hiver", kept "froid". The winter sense is a corpus artifact; both source rows cite 2 Timothy 4, which contains "come before winter" (v.21). CAR is equatorial - de is cold, not a season. - seleka: dropped "mariage", kept "alliance, pacte". Seleka means "alliance" in Sango. All three were duplicates introduced by a double corpus-import on 2026-05-21. CORRECTION to v1.4.0, which stated: "All 653 entries are unique by (sango, french); 4 Sango spellings recur with distinct French senses (legitimate polysemes, not duplicates)." Three of those four were NOT polysemes. Exactly one genuine polyseme remains - mu, glossed both "prendre, saisir" and "donner", with different argument structures (mu X vs mu X na Y). The card now says so and carries a dated correction. Counts recomputed from the file: 650 entries, 551 wit…adcaf77
-
Reconcile to 653 verified (computed from file); add provenance + DATASHEET; keep CC-BY-SA-4.0
Verified signaturef77104d