ghananlpcommunity/ghana-sentences

Ghana Sentences

A growing sentence-level text corpus for Ghanaian languages, tagged with ISO 639-3 codes and split into per-language subsets. The goal is broad-coverage text across all Ghanaian languages; this first release draws on school curriculum materials and a licensing-exam benchmark. More sources will be added over time.

Language list and ISO codes follow GhanaNLP/ghana-taught-local-languages.

Loading

from datasets import load_dataset

everything = load_dataset("ghananlpcommunity/ghana-sentences", "all", split="train")   # all languages
ewe        = load_dataset("ghananlpcommunity/ghana-sentences", "ewe", split="train")   # one language
akuapem    = load_dataset("ghananlpcommunity/ghana-sentences", "twi-aku", split="train")
asante     = load_dataset("ghananlpcommunity/ghana-sentences", "twi-asa", split="train")

Subsets

Subset	Language	ISO 639-3	Sentences	Words
ada	Dangme	ada	7707	91181
dag	Dagbani	dag	6451	64946
dga	Dagaare	dga	6065	62395
ewe	Ewe	ewe	5822	58132
fat	Fante	fat	584	5725
gaa	Ga	gaa	6160	59546
gjn	Gonja	gjn	6987	70892
gur	Gurene	gur	6243	70152
nzi	Nzema	nzi	7105	70770
twi-aku	Akuapem Twi	twi	7085	72631
twi-asa	Asante Twi	twi	6566	66841
xsm	Kasem	xsm	6781	78342

Total: 73556 sentences, 771553 words across 12 subsets.

Akuapem Twi and Asante Twi share ISO code twi, so they are shipped as two separate subsets — twi-aku (Akuapem) and twi-asa (Asante) — each independently deduplicated. The lang field is twi for both; subset and language_name carry the distinction.

Fields

• text — a single sentence in the local language
• subset — subset code (twi-aku, twi-asa, or the ISO 639-3 code)
• lang — ISO 639-3 code
• language_name — human-readable language / variety name
• source — origin of the sentence (curriculum or nsanku-mmlu)
• source_ref — original PDF filename or MMLU question id

Sources (first release)

• curriculum — learner + teacher materials from curriculumresources.edu.gh (233 PDFs), published by the National Council for Curriculum and Assessment (NaCCA), Ghana.
• nsanku-mmlu — questions and answer options from the nsanku Ghanaian-MMLU benchmark (NTC licensing-exam items).

How it was built (first release)

1. Curriculum PDFs downloaded; text extracted with PyMuPDF (orthography ɛ ɔ ŋ ɖ ƒ ɣ preserved). MMLU questions + answer options read from JSONL (option-letter prefixes like A. / b) stripped, person-name initials preserved).
2. Everything split into sentences; control characters, table-of-contents dot-leaders, and all-caps headings removed.
3. English and boilerplate sentences filtered out using an English-word-ratio heuristic (drop if ≥70% of tokens are common English words). Code-switched sentences (local main text with some English words) are kept.
4. Sentences deduplicated within each subset across both sources.

Caveats

• These are curriculum / exam texts; some English terms, phonetic notation, and code switching remain by design.
• Fante (fat) is under-represented — the curriculum portal currently publishes very little Fante; most Fante text here comes from the MMLU source.
• Copyright of the source materials rests with NaCCA / Ghana Ministry of Education and the NTC; this corpus is redistributed for research and language-technology development.

License

MIT (dataset packaging). Source content © NaCCA / NTC, Government of Ghana.