Datasets:

MEYNG
/

sango-vocabulary

Following the Datasheets for Datasets framework (Gebru et al., 2021, CACM) and informed by Data Statements for NLP (Bender & Friedman, 2018). All counts are computed directly from the published files — recounted 2026-09-04 from the shipped vocabulary.jsonl — and are honest about what the verified flag does and does not mean.

Dataset: MEYNG/sango-vocabulary — https://huggingface.co/datasets/MEYNG/sango-vocabulary
Version: 1.5.5 · Maintainer: MEYNG ([email protected])
_{Corrected 2026-09-05: this field read 1.4.7 while the metadata.json shipped
beside it read 1.5.4. The version is taken from that file, which is regenerated from
vocabulary.jsonl itself rather than written by hand. Some table captions in the dataset
card still say v1.5.0; those label when a specific breakdown was computed and are being
reviewed separately.}
Files: vocabulary.jsonl (590 entries), training_pairs.jsonl (360 pairs), grammar_rules.json (13 rules)
Languages: Sango (sag), French (fra), English (eng)
Licence: CC-BY-SA-4.0 — ⚠️ partly under review. Read Provenance and license status before redistributing.

Provenance and license status

The CC-BY-SA-4.0 claim over part of this dataset is under review. This section states the problem rather than leaving it to be discovered downstream.

316 of the 590 entries (53.6%) are NLLB-derived — nllb-enrichment (250), nllb_corpus (39), nllb_en_corpus (27) — extracted from the web-mined Sango–French bitext distributed by AllenAI. Counted directly from the shipped vocabulary.jsonl on 2026-09-04, not carried over from a previous release.

The upstream chain is ODC-BY, and ODC-BY carries an attribution requirement. That attribution is set out in the dataset card under Source attribution, and any redistributor must carry it forward. But ODC-BY governs AllenAI's compilation, not the crawled pages the bitext was mined from — the AllenAI card also binds users to "the respective Terms of Use and License of the original source". Those upstream terms are not established. Of the 70 entries carrying a source_url, 64 point at jw.org, a single publisher whose own terms do not permit redistribution.

That chain is recorded as CONTESTED in the SangoAI data-source registry (docs/SOURCES.md, § CONTESTED, verified 2026-09-04). MEYNG's internal data-governance policy forbids publishing the underlying 45,451-pair corpus on the compilation licence alone. That prohibition 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 identifying the upstream documents behind the Sango mine. That one check settles the question for this dataset and for 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.

If you need a licence chain clean to a named source today, filter source to the 264 entries with a named non-NLLB origin. The 6 admin and 4 blank rows are not a clean subset — their origin is unrecorded, which is a separate problem that avoiding NLLB does not solve.

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.

1. Motivation

Why was the dataset created? Sango is the co-official language of the Central African Republic (with French; ~5M speakers) and is severely under-represented in NLP. No open, structured, machine-readable trilingual Sango lexicon existed. This dataset is the seed lexical resource for SangoAI's vocabulary-augmented prompting and for any researcher/NGO building Sango NLP.
Who created it and for whom? MEYNG, an African-language-AI project founded by a native Sango speaker from the CAR, for the open low-resource-NLP community.
Who funded it? MEYNG (self-funded); AI enrichment used cloud LLM inference credits.

2. Composition

What do instances represent? Two kinds:
1. Vocabulary (590 rows): a Sango word/phrase with French + English translations, a semantic category, a difficulty level, optional example sentences, a pipeline confidence score (0–1, present on 100% of rows), a status, a source provenance tag, and source_pages — the page(s) of the CAR primer the headword came from, present on 210 of the 590 rows (new in v1.5.5). An empty source_pages means not established, never "no page": 46 primer-sourced rows matched no OCR worksheet row, and the 334 non-primer rows do not come from a book. Where a headword is taught on several pages, all of them are listed — 13 rows — because choosing one would be a guess presented as provenance.
2. Training pairs (360 rows): Sango→French text pairs extracted from a grammar/ dictionary reference.
How many instances? 590 vocabulary entries; 360 translation pairs; 13 grammar rules. Vocabulary is unique by (sango, french); 4 Sango spellings recur with distinct French senses (polysemes, deliberately kept — not duplicates).
Is it a sample or complete? It is the current verified subset of a larger, growing curation pipeline (4,995 additional candidates await native-speaker review).

What is the "verified" label? status: verified is a curation-pipeline flag (cross-referenced against textbook/dictionary/corpus sources). It is not native-speaker sign-off. Provenance of the 590 entries:

Provenance group	Entries
NLLB parallel-corpus enrichment	316
Textbook (CAR primer / dictionary)	256
Origin not recorded (`admin`)	6
Unlabelled (blank `source`)	4
Swadesh list	3
Native-speaker audit session (2026-04-20)	2
Native speaker (individually confirmed)	1
Wiktionary	1
Community contributor	1
Total	590

⚠️ Corrected 2026-09-04. The previous version of this table carried an "AI prompt enrichment — 18" row and counted unrecorded origins twice (6 + 6), summing to 610 against a 590-row file. The 18 AI-prompt-sourced entries were withdrawn on 2026-09-04 — MEYNG's data-honesty rule forbids generated Sango entering the corpus as attested, and those rows had been published as verified. Unrecorded origins are 6 admin + 4 blank =

Every figure above is recounted from the shipped file.

Are there labels/targets? The French/English translations are the targets for MT; category/difficulty support classification and curriculum tasks.
Missing information? Example sentences cover 81.7% (Sango, 482 of 590) and 67.3% (French, 397 of 590); English examples are absent. (Corrected 2026-09-04: this line read "~84.8% / ~68.1%" alongside its own "482 of 590", which is 81.7% — the percentages were stale against the counts beside them. Both recounted from the shipped file.) No pronunciation/audio field (0/590) — a known gap for a tonal language. Some training_pairs.jsonl rows are fragments, not clean sentences.
Confidential / sensitive / personal data? None. The data is general-domain vocabulary; it contains no personal, identifying, or sensitive information.
Dialect / register scope? Primarily standardised Sango as used in Bangui. Regional variants and specialised registers (medical, legal, agricultural) are under-represented.

3. Collection process

How was the data acquired? (a) OCR of the Kîrîndönî CAR primary-school Sango primer (photo-captured pages, cleaned to CSV); (b) extraction/enrichment from the quality-filtered NLLB fra_Latn–sag_Latn / eng_Latn–sag_Latn corpora (LASER ≥ 1.0, target-LID ≥ 0.9); (c) AI (LLM) enrichment of candidate translations; (d) a Swadesh list and French Wiktionary; (e) admin/manual entry.
japprendslesango.com is NOT an acquisition source. An earlier version of this list named it. No entry carries it as a source or source_url — checked across all 607 entries on 2026-08-13 — and its terms reserve all rights ("toute reproduction sans autorisation est interdite", conditions-generales, last updated 2025-02-24). It is credited in the dataset card's Acknowledgments as a resource that informed this work, which is what it was. Do not re-add it here without evidence of an entry that came from it.
Who was involved? MEYNG's founder (native Sango speaker) structured, categorised, and reviewed entries; automated pipelines performed extraction and enrichment.
Over what timeframe? Curated 2026-04 onward; this snapshot is v1.4.1 (2026-07).
Ethical review? No formal IRB; the data is non-personal public-domain vocabulary and educational material.

4. Preprocessing / cleaning / labelling

What was done? Deduplication (capitalisation variants and wrong-spelling collisions removed; polysemes preserved on the (sango, french) key); orthographic normalisation toward the standardised CAR Sango orthography (Diki-Kidiri; seven phonemic vowels + tonal diacritics); trilingual alignment; category/difficulty labelling; per-entry confidence and source tagging.
Is raw data saved? The candidate pool and intermediate CSVs are retained in the SangoAI repository (scripts/data-collection/).
Is the cleaning software available? The pipeline scripts (prioritisation, enrichment, bulk import) are in the SangoAI repository.

5. Uses

Intended uses: seed lexicon for Sango NLP; grounding for vocabulary-augmented prompting; language-learning/education apps; language documentation; cross-lingual transfer experiments.
Not (yet) suitable for, without added native verification: high-stakes medical/legal/financial translation; clean-parallel-sentence MT training from training_pairs.jsonl without filtering.
Honest note on the companion model. A fine-tuned NMT model (MEYNG/nllb-sango-finetuned-600m) was trained on the NLLB Sango corpus. It improves in-distribution BLEU/chrF but regresses below its untrained baseline on out-of-distribution (FLORES-200) text — it memorised a corpus domain rather than learning the language. Users should treat that model as domain-specific and prefer vocabulary-grounded prompting for out-of-domain and culturally-specific text.
What users should know to avoid harm: do not present pipeline-verified entries as native-speaker-certified; apply confidence filtering; expect Bangui-standard register; the 1-native-confirmed count is stated so reliability is not overstated.

6. Distribution

How distributed? HuggingFace Datasets Hub (MEYNG/sango-vocabulary), loadable via datasets.load_dataset.
Licence: CC-BY-SA-4.0 (attribution + share-alike). Chosen deliberately so derivatives stay open — the verified lexicon cannot be enclosed in a closed product. (The companion model is CC-BY-NC-4.0; the licences differ by design — the model is non-commercial, the data is open share-alike.)
Fees / restrictions: none beyond the licence.

7. Maintenance

Who maintains it? MEYNG ([email protected]).
How is it updated? The verified count grows as native-speaker verification proceeds (target: 3,000 verified entries). Updates are pushed to the Hub with a version bump and a refreshed card + this datasheet.
How can others contribute? Native Sango speakers, Ubangian-language linguists, and researchers can open PRs on the Hub; domain-specific additions (medical, legal, agricultural, financial) are specifically welcomed.
Erratum / versioning policy: counts on the card and in metadata.json are recomputed from the shipped file at each release; the live SangoAI API count is a separate, growing metric and may differ from any file snapshot.

Prepared 2026-07-07. If you find an error — especially a translation a native speaker would correct — please open an issue or PR; correcting entries toward native-speaker sign-off is the dataset's active roadmap.