This dataset addresses critical gaps in existing Gĩkũyũ (Kikuyu) NLP resources. Gĩkũyũ is the most widely spoken indigenous language in Kenya (approximately 8 million native speakers), yet it is absent from major African NLP benchmarks such as MasakhaNER and MasakhaPOS, and scores poorly in Meta's NLLB/FLORES-200 evaluation (3–9 BLEU).

The corpus was created by a native Gĩkũyũ speaker with clinical and NLP annotation expertise. Special attention was given to linguistic features that consistently cause errors in current models: diacritic accuracy (ĩ, ũ, â), the Nĩ vs Ndĩ copula distinction, affirmative and negative verb forms, tonal constructions, future and continuous tense prefixes, and natural modern register.

Key features

100+ human-verified parallel sentences (English ↔ Gĩkũyũ)
Expert-authored: no machine translation, no AI assistance
Correct orthography with full diacritic coverage
10 thematic domains: greetings, agriculture, health, education, family, technology, government, environment, emotions, complex syntax
Explicit notes on common model errors and culturally-bound expressions
Contributed to improving Grok (xAI) Gĩkũyũ language capabilities

Example data

English	Gĩkũyũ	Notes
I will listen attentively.	Niî nîngûgûthikîrîria wega.	Future prefix `nîngû-`
I will not listen.	Ndi gûthikîrîria.	Negative future; `Ndi` ≠ `Ndĩ` (copula)
You are a good student.	Wî mûrutwo mwega.	`Wî` = 2sg copula
I am here.	Niî ndî haha.	`Ndî` = 1sg copula; `Niî` = emphatic pronoun
Take this medicine twice a day.	Nyua dawa ĩno maita meerĩ oro mũthenya.	Medical domain
The rains came early this year.	Mbura nĩ ĩraurire tene mwaka ũyũ.	Agricultural domain

Linguistic notes

The `Nĩ` / `Ndĩ` distinction

This is one of the most common sources of model error in Gĩkũyũ:

Form	Function	Example
`Nĩ`	Copula (3rd person / equative)	Nĩ mwarimu — He/she is a teacher
`Ndĩ`	Copula (1st person singular)	Ndĩ mwarimu — I am a teacher
`Wî`	Copula (2nd person singular)	Wî mwarimu — You are a teacher

Diacritics

Gĩkũyũ uses circumflex and tilde diacritics that encode phonemic distinctions. Omitting them changes meaning. This corpus uses the standard orthography endorsed by the Gĩkũyũ Literacy Committee:

ĩ — nasalised /i/
ũ — nasalised /u/
â, î, û — long vowels in certain dialects

Dataset structure

gikuyu-english-parallel/
├── README.md                  ← this file
├── data/
│   ├── train.jsonl            ← main parallel data (JSONL format)
│   ├── train.csv              ← same data in CSV format
│   └── test.csv               ← held-out evaluation split (if available)
└── configs/
    └── label_studio_ner.xml   ← Label Studio annotation config

Column schema

Column	Type	Description
`id`	int	Unique sentence identifier
`english`	string	English source sentence
`gikuyu`	string	Gĩkũyũ target translation
`domain`	string	Thematic domain (health, greetings, etc.)
`notes`	string	Optional translator notes or correction flags

JSONL format (train.jsonl)

{"id": 1, "english": "Good morning, how did you sleep?", "gikuyu": "...", "domain": "greetings", "notes": ""}
{"id": 21, "english": "Take this medicine twice a day after eating.", "gikuyu": "...", "domain": "health", "notes": ""}

Domain coverage

Domain	Sentence IDs	NLP gap addressed
Greetings & Social	1–10	Core pragmatic patterns
Agriculture & Food	11–20	Rural subsistence vocabulary
Health & Medicine	21–32	Clinical NLP; zero prior Gĩkũyũ medical corpora
Education	33–42	School-domain vocabulary
Family & Relationships	43–52	Kinship terminology
Technology & Digital	53–62	Entirely absent from all prior Gĩkũyũ NLP data
Government & Civic	63–70	Civic literacy language
Environment & Nature	71–78	Ecological and geographic vocabulary
Emotions & Mental States	79–86	Sentiment and affect data
Complex Syntax	87–100	Conditionals, interrogatives, subordinate clauses

Intended uses

Recommended:

Machine translation training and evaluation (English ↔ Gĩkũyũ)
Fine-tuning multilingual LLMs (mBART, NLLB, M2M-100) for Gĩkũyũ
Spelling and diacritic correction tools
Speech synthesis and ASR alignment for Gĩkũyũ
Cultural preservation and language education

Not recommended:

High-stakes automated decision-making without human review
Standalone production MT without additional training data

Languages

Language	BCP-47	ISO 639-3	Notes
English	en	eng	Standard written English
Gĩkũyũ	kik	kik	Central Kenya dialect; modern register

Comparison with existing resources

Resource	Type	Gĩkũyũ quality	Domain coverage
Meta NLLB (FLORES-200)	Machine translation	BLEU 3–9 (poor)	General
MasakhaNER	NER benchmark	Not included	News
Khaya AI	MT model	BLEU ~16	General
This dataset	Human parallel corpus	Native speaker	10 domains

Creator

Shadrack Mwangi
Pharmacist & NLP Data Specialist · KC Children's Hospital, Nairobi, Kenya
Native Gĩkũyũ speaker · AI/NLP annotator (medical and African languages)

Contributed to Grok (xAI) Gĩkũyũ language training data

HuggingFace: Mwangi86
GitHub:
LinkedIn / linkedin.com/shadrack-mwangi-243a66349

License

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

You are free to share and adapt the material for any purpose, including commercially, provided you give appropriate credit.

Citation

@misc{mwangi2026_gikuyu_parallel,
  title     = {High-Quality {E}nglish--{G}{\~{i}}k{\~{u}}y{\~{u}} Parallel Corpus},
  author    = {Mwangi, Shadrack},
  year      = {2026},
  publisher = {HuggingFace},
  url       = {https://huggingface.co/datasets/mwangi86/gikuyu-english-parallel},
  note      = {Human-curated parallel corpus for low-resource Gĩkũyũ NLP.
               Native speaker translations covering 10 thematic domains.}
}

Related datasets

SwahiliMed — Swahili–English medical parallel corpus by the same author
FLORES-200 — Meta multilingual benchmark (includes kik, low quality)
CGIAR/KikuyuEnglish_translation — Agriculture-domain Gĩkũyũ data

Changelog

Version	Date	Notes
v1.0	June 2026	Initial release — 100 sentences, 10 domains

Downloads last month: 27

Total file size:

44.2 kB

You need to agree to share your contact information to access this dataset

English–Gĩkũyũ Parallel Corpus (Human-Curated)

Dataset summary