MEYNG/sango-vocabulary

Sango Vocabulary Dataset

Dataset Description

An open, structured, machine-readable trilingual vocabulary dataset for Sango (ISO 639-1: sg, ISO 639-3: sag), the co-official language of the Central African Republic (with French) and its most widely spoken language. Sango is a creole language with over 5 million speakers, yet it remains severely underrepresented in NLP research and digital resources.

This dataset provides trilingual vocabulary entries (Sango-French-English) with semantic categorization and example sentences, along with Sango-French translation pairs extracted from grammar and dictionary references.

Supported Tasks

• Machine Translation: Sango-French and Sango-English translation
• Language Modeling: Pre-training or fine-tuning language models for Sango
• Cross-lingual Transfer: Leveraging French/English representations for Sango NLP
• Language Documentation: Digital preservation of Sango vocabulary and usage patterns
• Educational Technology: Building language learning applications for Sango

Languages

Language	ISO 639-1	ISO 639-3	Role
Sango	sg	sag	Primary target language
French	fr	fra	Primary bridge language
English	en	eng	Secondary bridge language

Dataset Structure

Vocabulary (vocabulary.jsonl)

590 verified trilingual vocabulary entries (v1.5.0), each containing:

Field	Type	Description	Coverage
word_id	string	Stable UUID for the entry	100%
sango	string	Sango word or phrase	100%
french	string	French translation	100%
english	string	English translation	100%
category	string	Semantic / grammatical category	100%
difficulty	string	Learning level (beginner/intermediate/advanced/expert)	100%
example_sango	string	Example sentence in Sango	~84%
example_french	string	Example sentence in French	~67%
confidence	float	Curation-pipeline confidence score (0–1)	100%
status	string	Curation status (currently all verified)	100%
source	string	Provenance tag (e.g. textbook-myc, nllb-enrichment, admin)	~99%
source_url	string	Source URL where applicable	partial
source_pages	int[]	Page(s) of the CAR primer the headword was extracted from	210/590

Example entry:

{
  "word_id": "a4dec056-d560-4733-a8cf-47514bedb410",
  "sango": "abata",
  "french": "garder, préserver, maintenir",
  "english": "keep, preserve, maintain",
  "category": "verb_action",
  "difficulty": "intermediate",
  "example_sango": "Zo so abata atënë ayeke na kube.",
  "example_french": "Celui qui garde les paroles est béni.",
  "confidence": 0.72,
  "status": "verified",
  "source": "nllb_en_corpus"
}

Training Pairs (training_pairs.jsonl)

360 Sango-French translation pairs extracted from grammar and dictionary reference materials.

Field	Type	Description
source_lang	string	Source language code (sg or fr)
target_lang	string	Target language code (sg or fr)
source	string	Source text
target	string	Target text

Note on quality: These pairs were extracted from a Sango grammar reference book. While many entries contain genuine Sango-French translations and vocabulary definitions, some are sentence fragments or contextual excerpts from the source material. Users should apply filtering for downstream tasks that require clean parallel sentences.

Example entry:

{
  "source_lang": "sg",
  "target_lang": "fr",
  "source": "Mbï yê Bêafrîka mîngi",
  "target": "j'aime beaucoup la Centrafrique"
}

Category Distribution

The vocabulary spans 36 semantic categories. In v1.4.8 the nine singular/plural category-name duplicates left by earlier import batches (adjective/adjectives, animal/animals, emotion/emotions, number/numbers, place/places, profession/professions, and singular-only stragglers of color, greeting, verb in the source database) were merged into their plural canonical forms — no entries were added or removed. The largest (v1.4.8, computed from this file): essential (162), verbs (59), places (54), adjectives (30), numbers (28), actions (24), food (21), time (20), emotions (18), family and noun (17 each). For the exact, current per-category distribution, use the HuggingFace dataset viewer — counts shift as the dataset grows, so they are not hardcoded here.

Difficulty Distribution

Level	Count	Percentage
Beginner	338	51.8%
Intermediate	219	33.5%
Advanced	95	14.5%
Expert	1	0.2%

Dataset Creation

Source Data

• Vocabulary entries: Curated by MEYNG from the Kîrîndönî CAR primary-school Sango primer, a quality-filtered French-Sango parallel corpus (NLLB), a Swadesh list, French Wiktionary, and manual entry. Provenance is recorded per entry in the source field. No entry derives from japprendslesango.com — that site is credited in Acknowledgments as a resource that informed the work, not as a source of content (verified across all 607 entries, 2026-08-13).
• Training pairs: Extracted from a comprehensive Sango grammar and dictionary reference work.
• Verification: Entries are cross-referenced against the above sources and flagged verified in the curation pipeline.

Annotation Process

Vocabulary entries were structured and categorized by MEYNG's founder, a native Sango speaker, drawing on textbook, dictionary, and parallel-corpus sources. Each entry carries a confidence score and a status flag from the curation pipeline. Most entries are pipeline-verified (cross-referenced against the sources above) rather than individually confirmed by a native speaker — broader native-speaker re-verification, via a community panel, is in progress.

Provenance breakdown (v1.5.0, all 590 entries)

status: verified is a pipeline flag. The table below shows how each entry actually entered the set, so users can judge reliability directly rather than trusting a single "verified" label. Only 1 of 590 entries has so far been individually confirmed by a native speaker (source: native_speaker); the rest are textbook-, corpus-, or AI-derived and cross-referenced. Growing the native-speaker-confirmed share is the dataset's active roadmap (see the SangoAI vocabulary-verification workflow).

Provenance group	Entries	Origin
NLLB parallel-corpus enrichment	316	nllb-enrichment, nllb_corpus, nllb_en_corpus
Textbook (CAR primer / dictionary)	256	textbook-myc-review, textbook-batch-import, textbook-myc, textbook-myc-csv, textbook-myc-manual
Origin not recorded	6	admin
Native-speaker audit session (2026-04-20)	2	native-speaker-audit-2026-04-20 — not native-confirmed, see below
Community contributor	1	contributor
AI prompt enrichment	0	prompt-vocabulary — all 18 withdrawn 2026-09-04
Swadesh list	3	swadesh_list
Wiktionary	1	fr_wiktionary
Native speaker (individually confirmed)	1	native_speaker
Unlabelled	6	(blank source)

For the full question-by-question provenance and collection detail, see DATASHEET.md (Datasheet for Datasets, Gebru et al. 2021).

Considerations for Using the Data

Known Limitations

1. Dataset size: This release contains 590 verified vocabulary entries (v1.5.0) and 360 translation pairs, plus grammar_rules.json with 13 structured Sango grammar rules. The vocabulary was consolidated and deduplicated from a larger raw set (capitalization variants and wrong-spelling collisions removed). All 590 entries are unique by (sango, french); 1 Sango spelling (mû) recurs with distinct French senses — a genuine polyseme, glossed both "prendre, saisir" and "donner" with different argument structures (mû X vs mû X na Y). Corrected 2026-07-22: v1.4.0 stated 4 such spellings and called them all legitimate polysemes. Three were not — yongoro (same sense reworded), dê (a "winter" gloss lifted from 2 Timothy 4:21, not a Sango sense) and séléka (a "mariage" gloss; Séléka means alliance in Sango). All three were duplicates from a double corpus-import on 2026-05-21 and have been removed.

2. Verification semantics: status: verified is a curation-pipeline flag (cross-referenced against textbook, dictionary, and parallel-corpus sources). The majority of entries were machine-enriched or textbook-imported; only a small fraction have been individually confirmed by a native speaker so far. Treat confidence accordingly and apply your own filtering for high-stakes use.

3. Training pair quality: The translation pairs contain some fragmented text from the source material. Filtering is recommended for tasks requiring clean parallel data.

4. Example sentences: 482 of 590 entries (81.7%) include a Sango example sentence and 397 (67.3%) include a French example; English examples are not yet included. Coverage fell again in v1.4.9, also on purpose: 21 example sentences were removed because their provenance is the jw.org-dominated NLLB bitext (1,289 of 1,406 entries carrying a source_url point at that host) and one — for apendere — was recognisably Acts 2:11, i.e. text from a copyrighted Bible translation that cannot be relicensed under CC-BY-SA. The other 20 were short and almost certainly carried no protected expression, but shared the provenance and the register; a general-purpose lexicon whose examples skew religious is less useful for the health, education and administrative domains this dataset is meant to serve. Two of the removed sentences were additionally not Sango — "Kristo azali Yesu." and "Testament ya Yesu." use azali and ya, which are Lingala forms (Sango: ayeke, ti); this was spotted by tooling, not by a speaker, and awaits confirmation. The 21 headwords and their glosses are unchanged and remain published; only the example field was cleared, and all 21 are queued for native-speaker elicitation. Coverage had already fallen from 84.8% in v1.4.1 on purpose: 18 example sentences that did not contain their own headword were removed on 2026-07-22 rather than left in place — séléka was illustrated by a sentence not containing the word séléka, doli (éléphant) by one about a monkey. They are not being replaced with generated text: writing plausible Sango to fill the gap would inject invented language into a corpus whose value is that it documents a real one. Attested replacement candidates have been mined from the aligned corpus and await native-speaker approval. Every remaining example now contains its headword, apart from two multi-word entries matched discontinuously (correct Sango).

5. Pronunciation: The current release does not include a pronunciation field. Learner-oriented phonetic guides are available in the SangoAI app and are planned for a future dataset version.

6. Dialect coverage: Sango has regional variations. This dataset primarily represents the standardized Sango used in Bangui (the capital).

7. Gloss reliability (new in v1.4.2): every textbook-sourced gloss was checked against a 45,451-pair aligned FR–Sango corpus. 41 entries whose gloss was directly contradicted by attested usage were withdrawn — e.g. bîa was glossed couper while every corpus occurrence aligns with chanson/chante/musique; nzë was glossed moustique against mois/lune/année. They are moved to pending re-glossing, not deleted. 12 entries could not be corroborated OR refuted and remain published, individually identified in the project's TERMINAL-TRIAGE.csv. They are everyday concrete vocabulary — éléphant, pioche, piège — occurring 11–42 times in the aligned corpora against 1,721–18,531 for register-typical words. The corpus is structurally blind to this vocabulary class, so more of the same text would not settle them; they need a native speaker or a non-religious corpus. Treat those 12 as unverified rather than verified. A further 3 previously flagged entries were cleared by the corpus; the audit deliberately reports a review queue, not verdicts, because the method produces false positives (it flagged yeke=être, which is correct — French inflection defeats the matcher). Roughly a third of this dataset is textbook-sourced and had never been corpus-checked before v1.4.2. Extended in v1.4.7: the classifier behind that audit promoted an entry to "corroborated" on a single agreeing sentence, which removed 83 entries from review permanently. Re-running it with a rate threshold surfaced 45 entries agreeing on under 5% of their own attested occurrences, and two of them reproduced withdrawals already made on the same evidence: nze (sueur/transpirer, 0.0% in-context against lune/mois) and nzere (rire/sourire, 1.3% against plaire/agréable). Their spelling twins nzë and anzere had been withdrawn in earlier rounds while these were left published. Both are now withdrawn to pending, 609 → 607. The remaining 43 are a review sheet, not a queue of errors: several are synonyms the matcher cannot see (bâgara = vache/boeuf against taureaux) and one is a demonstrative particle with no competing sense at all.

Recommendations

• Use as a seed dataset to bootstrap Sango NLP systems
• Access the full dataset directly via the SangoAI API at https://sangoai.sbs (the live verified count grows over time and may exceed this file's snapshot)
• Apply quality filtering on training_pairs.jsonl (and on confidence) before fine-tuning
• Consider data augmentation techniques for downstream tasks

Citation

If you use this dataset in your research, please cite:

@dataset{meyng_sango_vocabulary_2026,
  title={Sango Vocabulary Dataset: A Trilingual Lexical Resource for an Underrepresented African Language},
  author={MEYNG},
  year={2026},
  url={https://huggingface.co/datasets/MEYNG/sango-vocabulary},
  license={CC-BY-SA-4.0},
  language={sg, fr, en}
}

License

This dataset is released under the Creative Commons Attribution-ShareAlike 4.0 International License (CC-BY-SA-4.0).

You are free to:

• Share -- copy and redistribute the material in any medium or format
• Adapt -- remix, transform, and build upon the material for any purpose, including commercial

Under the following terms:

• Attribution -- You must give appropriate credit to MEYNG, provide a link to the license, and indicate if changes were made
• ShareAlike -- If you remix, transform, or build upon the material, you must distribute your contributions under the same license

What the CC-BY-SA-4.0 grant covers

This dataset is curated, not wholly original. The CC-BY-SA-4.0 grant covers MEYNG's curation work -- the selection, structuring, categorisation, trilingual alignment, tone-diacritic normalisation and verification pipeline, plus the entries authored outright. It is not a claim of authorship over the underlying sources.

⚠️ Provenance and license status — the claim over the NLLB-derived half is under review

316 of the 590 entries (53.6%) are NLLB-derived, and the CC-BY-SA-4.0 claim over that half is under review. It is stated here rather than left to be discovered downstream.

The NLLB bitext reaches us under ODC-BY, whose attribution requirement is set out in Source attribution immediately below. ODC-BY governs AllenAI's compilation, not the crawled pages the bitext was mined from, and the AllenAI card binds users to "the respective Terms of Use and License of the original source". Those upstream terms are not established — and of the 70 entries carrying a source_url, 64 point at jw.org, a publisher whose own terms do not permit redistribution.

MEYNG records this chain as CONTESTED in its data-source registry (docs/SOURCES.md, § CONTESTED, verified 2026-09-04), and its governing document forbids publishing the underlying 45,451-pair corpus on the compilation licence alone. That rule was written for the corpus and does not retroactively cover this derived lexicon, which was already public when the chain was marked CONTESTED. We are naming that gap rather than sheltering behind it.

What resolves it: whether AllenAI publishes per-language provenance for sag_Latn naming the upstream documents behind the Sango mine. That single check settles the question for this dataset and the corpus together.

What we commit to once it resolves:

Outcome	What we do
Upstream permissive	Keep the dataset at full size and carry the attribution chain explicitly, naming the upstream terms alongside ODC-BY.
Upstream restricted	Cut a clean release from the 274 non-NLLB entries and withhold the 316 NLLB-derived rows until they can be re-derived from CLEARED sources.
Outcome ambiguous	Take IP counsel before choosing between the two. We will not settle an ambiguous licence question by picking whichever answer keeps the dataset larger.

Until it resolves the dataset stays published at full size with this disclosure attached. That is a deliberate decision rather than an oversight, and the restricted case remains reversible via the clean-cut release above. The same statement in datasheet form is in DATASHEET.md § Provenance and license status.

Source attribution (required in addition to the above)

316 of the 590 vocabulary entries derive from the NLLB web-mined Sango-French parallel corpus, via quality-filtered extraction (LASER >= 1.0, target-language-ID >= 0.9). That corpus is distributed by AllenAI under ODC-BY, which requires attribution. If you redistribute or adapt this dataset, carry that credit forward alongside MEYNG's:

Contains material derived from the NLLB web-mined bitext (AllenAI distribution, ODC-BY). NLLB Team et al., No Language Left Behind: Scaling Human-Centered Machine Translation, arXiv:2207.04672 (2022).

Two clarifications, because these are routinely confused:

• ODC-BY is attribution-only. It does not restrict commercial use, and it does not conflict with CC-BY-SA-4.0 downstream. The credit is mandatory; the permission is not narrowed.
• This is a different licence chain from the NLLB model weights. Those are CC-BY-NC-4.0 and non-commercial. The bitext is not. Do not carry the NC term across from one to the other.

The AllenAI card also states that users are "bound to the respective Terms of Use and License of the original source" -- the NLLB bitext is web-mined, so individual sentences carry whatever terms their origin site imposes. ODC-BY governs the compilation, not every underlying page.

A further 256 entries derive from a Central African Republic language primer / dictionary reference (the MYC / Kîrîndönî primary-school Sango primer). That work has not been identified precisely enough to state its terms or write an attribution for it; we are recording the gap rather than implying it does not exist.

New in v1.5.5 — source_pages. Those entries now carry the page(s) of that primer they were extracted from, recovered from the original OCR worksheets. 210 of the 256 carry at least one page; 197 resolve to a single page and 13 to more than one, and those keep every page — a headword taught on p.1 and revisited on p.55 is on both, and picking one would be a guess presented as provenance. The remaining 46 matched no worksheet row and carry [].

source_pages: [] means "not established", never "no page". The 334 non-primer entries carry [] because they do not come from a book at all. Do not read an empty list as a negative finding.

This does not resolve the licence question — it makes it answerable. Until now the primer could only be described; each of those 210 entries can now be pointed at a specific page of a specific book, which is what an attribution and a permission request both require.

Where the material actually comes from

Counted directly from the published vocabulary.jsonl at v1.5.0 (2026-09-04):

Provenance (source field)	Entries
NLLB web-mined bitext (nllb-enrichment, nllb_corpus, nllb_en_corpus)	316
Named non-NLLB sources (CAR primer 256, Swadesh 3, audit session 2, French Wiktionary 1, native speaker 1, community contributor 1)	264
admin — origin not recorded beyond the import route	6
prompt-vocabulary — AI-prompt sourced	0 (all 18 withdrawn 2026-09-04)
no source value	4
Total	590

(Corrected 2026-09-04: the blank-source row read 6, which made this table sum to 592 against a 590-row file. Recounted from the shipped vocabulary.jsonl: 4 rows carry an empty source. The "10 admin / unlabelled entries" figure below is unaffected — 6 + 4 = 10.)

The dominant upstream host is jw.org. Of the 70 entries carrying a source_url, 64 point there — a single publisher whose own terms do not permit redistribution. (A larger figure, 1,289 of 1,406, appears under Limitations above; that describes the mining pool entries were selected from, not these 590 published rows. They are different denominators and should not be quoted interchangeably.)

We do not think this makes the vocabulary entries unusable, and the reasoning belongs in the open rather than left implied:

• A headword and its gloss are facts about a language, not authored expression. The 316 NLLB-derived entries are lexical facts extracted from aligned sentences.
• Example sentences are a different matter, because a sentence can carry protected expression. That is why 21 were removed in v1.4.9 — one was recognisably Acts 2:11 from a copyrighted Bible translation — and why replacements are elicited from native speakers rather than mined.

This is our reasoning, not legal advice, and one question is genuinely open: the EU sui generis database right protects substantial extraction from a database independently of whether its contents are facts. We have not taken advice on whether it applies here.

If you need a licence chain clean to a named source, filter source to the 264 entries in the "named non-NLLB" row above. The 10 admin / unlabelled entries are not a clean subset — their provenance is unrecorded or generated, which is a separate problem from the NLLB chain and is not solved by avoiding it.

native-speaker-audit-2026-04-20 is not a native-speaker confirmation

Two entries carry that source. They were created during a native-speaker audit session on 2026-04-20 — a real, dated event — but that is not the same as having been confirmed by a speaker, and they are counted in their own provenance bucket rather than folded into native_speaker_confirmed.

The evidence for keeping them separate: that same audit updated 19 other entries and changed the source on none of them (14 primer, 3 NLLB-derived, 2 since withdrawn). If the convention were "audited by a speaker ⇒ native-sourced", those 19 would say so. And the entries that do read native_speaker were created a month later by an unrelated process.

So the native-speaker-confirmed count remains 1. Folding these in would have made it 3 on the strength of a session name.

Known defect: status does not carry the verification tier

Every one of the 590 rows reads status: verified. "Verified" here means the curation pipeline accepted the entry. (The 18 AI-sourced entries that previously carried this flag were withdrawn on 2026-09-04 — see the provenance table.) It does not mean a native speaker confirmed it — exactly one entry has that, as stated under Limitations. The intended three-tier model (candidate / pipeline_verified / native_verified) is not yet represented in this file. Until it is, read status as "present in the curated set" and nothing stronger.

Acknowledgments

This dataset was created by MEYNG as part of the SangoAI project, an AI-powered language platform dedicated to the preservation and digitization of Sango.

We acknowledge:

• The people of the Central African Republic, whose language and culture this dataset aims to preserve and promote
• The Sango-speaking community worldwide for their contributions to language documentation
• The creators of japprendslesango.com and other Sango learning resources, whose work informed ours. No content from those sites is reproduced in this dataset — japprendslesango.com reserves all rights, and we respect that
• Meta AI and AllenAI, for the NLLB project and the web-mined bitext that 316 of these entries derive from (see Source attribution above)
• The open-source NLP community working on low-resource African languages

Contact

• Organization: MEYNG
• Website: meyng.com
• Platform: sangoai.sbs
• Email: contact@meyng.com
• GitHub: github.com/meyng-hub/sangoai