Taimwe/kiribati-language-dataset

Kiribati (Gilbertese) Language AI Learning Dataset

A structured, source-cited dataset for learning/teaching the Kiribati language (taetae ni Kiribati, ISO 639: gil / ki), arranged from the alphabet → phonology → grammar → numbers → vocabulary → example sentences / use cases. Built for AI tutoring, translation pairs, and fine-tuning.

🔒 Access control: this repository is gated with manual approval ("Manually approve" setting). Only users you personally approve can see/download the data. Everyone else must request access, and the decision is yours. The data is additionally protected by the custom Permission License in LICENSE: redistribution, inclusion in other public datasets, model training/publication, and commercial use all require your prior written permission.

Data

Single split, train, as JSONL (one record per line): data/train.jsonl — 279 records, 16 task types.

Field	Type	Description
id	string	Stable record ID (kir-00001…)
task	string	Record type (see below)
category	string	Curriculum section it came from
kiribati	string	Kiribati text
english	string	English translation/definition
ipa	string	IPA (alphabet records)
gloss	string	Word-by-word literal gloss (example sentences)
context	string	Situation (phrases)
source	string	Cited source(s)
confidence	string	verified (in fetched sources) / common-unverified (verify before training)
notes	string	Extra info

Task types & counts

word 125 · number 30 · phrase 30 · alphabet 23 · example_sentence 17 · interrogative 16 · pronoun 6 · verb_particle 6 · number_old_system 6 · numeral_classifier 8 · demonstrative 4 · adverbial_demonstrative 3 · reduplication 3 · article 1 · sample_text 1

Loading

from datasets import load_dataset
ds = load_dataset("YOUR_USERNAME/kiribati-language-dataset", split="train")
print(ds[0])  # {'id': 'kir-00001', 'task': 'alphabet', 'letter': 'A a', 'ipa': '/a/', 'kiribati': 'ataei', 'english': 'child', ...}

Raw JSONL also works with no dependencies:

import json
rows = [json.loads(l) for l in open("data/train.jsonl", encoding="utf-8")]

Curriculum (markdown study files)

File	Content
01_overview_alphabet_phonology.md	Overview, dialects, alphabet (23 letters, IPA, examples), phonology
02_grammar.md	Word order (VOS), articles, nouns/classes, pronoun tables, possession, demonstratives, adjectives, verbs & particles, reduplication, negation, copula/existential, conjunctions, prepositions, interrogatives
03_numbers.md	Numerals + classifier system, ordinals, fractions, old native large-number system
04_vocabulary.md	~125 curated core words in 11 themes with confidence tags
05_examples_usecases.md	Glossed example sentences, ~30 practical phrases, sample text, AI use cases
build_dataset.py	Reproducibly regenerates data/train.jsonl from the markdown files
SOURCES.md	Full source list + licensing notes

Method & provenance

• Compiled from online research (Wikipedia CC BY-SA, Omniglot, Trussel.com/Peace Corps handbook via Wayback Machine, Internet Archive, ERIC). Every record carries its source.
• Included: factual linguistic data — orthography, IPA inventory, grammar rules, pronoun paradigms, numerals, word–translation pairs, short phrases/sentences — re-organized and re-written with citation.
• Intentionally NOT dumped verbatim: full texts of copyrighted references (Sabatier's dictionary, Trussel/Peace Corps handbook, Trussel–Groves dictionary). Bingham (1908) is public domain and is the recommended target for bulk expansion. See SOURCES.md.

Suggested AI uses

1. Tutoring agent following the curriculum order (alphabet → … → use cases).
2. Translation/eval from the phrase and example_sentence pairs.
3. Grammar drills: negation (aki), reduplication (habitual/continuative), passive (-aki), causative (ka-V-a), numerals with classifiers.

Gaps / next steps

• Harvest the public-domain Bingham (1908) dictionary (~179 pp) for a full wordlist.
• Add audio (Kaipuleohone archive, Univ. of Hawai'i).
• Verify common-unverified items & NGG examples against Sabatier/Bingham.

Compiled 2026-10-15. The Kiribati Language Board is the normative authority for the standard language.