CIL-Dataset: Colombian Indigenous Languages

Parallel and monolingual corpora for machine translation between Spanish and four Colombian Indigenous languages: Wayuunaiki, Inga, Kamëntsá and Nasa Yuwe. Compiled for the thesis Data-Centric Strategies for Low-Resource Neural Machine Translation in Colombian Indigenous Languages (Universidad de los Andes, FLAG Lab).

Languages

Language	Column	ISO 639-3
Spanish	esp	spa
Wayuunaiki	way	guc
Inga	ing	inb
Kamëntsá	kam	kbh
Nasa Yuwe	nas	pbb

Contents

Parallel and monolingual corpora

Full corpora, one aligned segment per row.

Config	File	Rows	Columns
inga	inga.csv	4,552	esp, ing, origen
kamentsa	kamentsa.csv	281	esp, kam, origen
nasa_yuwe	nasa_yuwe.csv	4,031	esp, nas, origen
wayuunaiki	wayuunaiki.csv	83,579	esp, way, origen
wayuunaiki_monolingual	wayuunaiki_monolingual.csv	2,676	way, origen

The origen column is a provenance identifier formatted as <language>__<type>_<source>__<year> (for example wayuu__dictionary_negrete_amaya__2021). Sources include Bible translations, dictionaries, grammars, constitutional texts, short stories and books.

Rephrasing subset (Wayuunaiki)

Train/validation/test splits used to study rogue memorization. In the training split, the Spanish side of biblical rows is replaced by an alternate human Spanish Bible translation, which reduces repeated Spanish targets. The Wayuunaiki side is never modified, and validation/test keep the original Spanish.

Config	Split	Rows
wayuunaiki_rephrasing_completo	train / validation / test	67,135 / 8,222 / 8,222
wayuunaiki_rephrasing_sin_dict	train / validation / test	7,469 / 763 / 764

Extra columns in the rephrasing files: source_row_id, esp_original, esp_target_rephrased, verse_id, repeat_group_size, assigned_version, rephrase_applied, rephrase_strategy.

Usage

from datasets import load_dataset

# a parallel corpus
inga = load_dataset("Flaglab/CIL-Dataset", "inga")

# the rephrasing splits
reph = load_dataset("Flaglab/CIL-Dataset", "wayuunaiki_rephrasing_completo")
print(reph["train"][0])

Trained models

NLLB-200 fine-tunes built on this data are available in the collection Machine Translation for Colombian Indigenous Languages.

Sources and licensing

The corpora were assembled from heterogeneous materials (Bible translations, dictionaries, grammars, constitutional texts, community stories and books). These sources carry their own copyrights and licenses. The dataset is released for non-commercial research use; downstream users are responsible for respecting the licensing of each underlying source. The rephrasing subset draws Spanish renderings from public Spanish Bible versions in the BibleNLP/eBible corpus.