We developed and curated ai4ug/luganda_prose as a monolingual Luganda text corpus to support language model pretraining, continual pretraining (CPT), morpheme-aware tokenizer development, and linguistic research. Version 2.0.0 contains 998,737 records comprising 32,320,315 words across sentences, paragraphs, news articles, encyclopedic passages, literary text, and vocabulary entries.

We built this corpus at the AI for Uganda Initiative in support of Pearl Chat, a native Luganda conversational language model project. Our goal is to provide a clean, large-scale collection of Luganda text for researchers and developers building language technologies for African languages.

Each record contains a text field and associated metadata describing its content type, subject domain, language, script, quality score, text length, and assigned dataset split.

Dataset summary

Property	Value
Records	998,737
Total words	32,320,315
Language	Luganda
ISO 639-3	`lug`
Script	Latin (`Latn`)
Text normalization	Unicode NFC
Corpus type	Monolingual text
Primary applications	Language model training, tokenizer development, corpus research
Splits	Train, validation, test
Split allocation	80%, 10%, 10%
Author & Maintainer	Wesley Kambale (AI for Uganda Initiative)
Version	`2.0.0`
License	Apache-2.0

Intended uses

We designed this corpus to support research and development in Luganda language modelling and natural language processing:

Continual pretraining: Adapting language models to Luganda vocabulary, grammatical structures, and textual patterns.
Autoregressive language modelling: Training or adapting models to predict subsequent tokens in natural Luganda text.
Tokenizer development: Training and evaluating subword and morpheme-aware tokenizers for Luganda and related Bantu languages.
Masked language modelling: Supporting representation learning and masked-token prediction experiments.
Linguistic research: Studying lexical distributions, sentence structures, noun-class concords, and register variation across Luganda text types.
African language NLP: Advancing foundational language technologies for Luganda.

Dataset composition

Subject-domain distribution

We classified all records across subject domains covering news, society, governance, agriculture, geography, health, business, education, family, and culture:

Domain	Records	Share
Grammar and news	596,206	72.83%
General society	48,757	5.96%
Religion and literature	30,755	3.76%
Society and news	30,000	3.66%
Encyclopedic knowledge	21,114	2.58%
Civic life and law	16,200	1.98%
Agriculture	15,949	1.95%
Geography and tourism	10,288	1.26%
Health and medicine	8,385	1.02%
Business and finance	6,723	0.82%
History, culture and kingdoms	6,467	0.79%
Education and literacy	6,360	0.78%
Family, parenting and youth	6,140	0.75%
Sports and entertainment	3,154	0.39%
Health and maternal care	2,805	0.34%
Urban life and transport	2,638	0.32%
Science, medicine and artisans	2,514	0.31%
Science and technology	1,986	0.24%
Vocabulary and grammar	770	0.09%
Grammar and linguistics	619	0.08%
News and society	500	0.06%
Culture and society	270	0.03%
Spoken conversation	45	0.01%
Total	818,645	100%

Content-type distribution

The type field describes the structural form of each record:

Content type	Records	Share
`sentence`	680,961	83.18%
`news_paragraph`	50,062	6.12%
`paragraph`	43,885	5.36%
`encyclopedic_paragraph`	21,114	2.58%
`news_article`	18,923	2.31%
`vocabulary_term`	2,430	0.30%
`vocabulary_definition`	770	0.09%
`literary_paragraph`	445	0.05%
`spoken_utterance`	45	0.01%
Passages and grammar explanations	10	<0.01%
Total	818,645	100%

Dataset splits and deduplication

We partitioned the corpus into training, validation, and test splits using an 80/10/10 allocation:

Split	Records	Share	Words
Train	835,745	83.7%	28,499,997
Validation	81,380	8.1%	1,889,295
Test	81,612	8.2%	1,931,023
Total	998,737	100%	32,320,315

Split assignment

In our split construction pipeline, records sharing the same six-word normalized text prefix are deterministically hashed into the same partition so closely related sentences do not straddle training and evaluation splits.

Our verification checks confirm:

Exact text overlap across the three splits: 0.0%.
Exact cross-dataset overlap with ai4ug/luganda_qa_pairs: 0.0%.

Data format and schema

Each record contains a text sample and associated metadata:

Field	Type	Description
`id`	String	Unique record identifier
`text`	String	Luganda sentence, paragraph, article, or lexical entry
`type`	String	Content-type classification
`split`	String	Dataset partition (`train`, `validation`, `test`)
`added`	String	Record date in `YYYY-MM-DD` format
`domain`	String	Subject-domain classification
`language`	String	Language code, `lug`
`script`	String	Script code, `Latn`
`quality_score`	Number	Quality score (`>= 0.80`)
`char_length`	Integer	Character count
`word_count`	Integer	Word count
`metadata`	Object	Structured record-level metadata

Example record

{
  "id": "lug_cpt_0000001",
  "text": "Okusobola okukakasa amakungula amalungi, kikulu okulowooza ku mbeera y'obudde esinga obulungi ey'okusimba. Abalimi b'ensuku bangi bakizuula nti enva endiirwa ezimu zikula bulungi bwe zibeera mu bbugumu ery'amaanyi.",
  "type": "paragraph",
  "split": "train",
  "added": "2026-10-15",
  "domain": "general_society",
  "language": "lug",
  "script": "Latn",
  "quality_score": 0.94,
  "char_length": 216,
  "word_count": 29,
  "metadata": {
    "language": "lug",
    "script": "Latn",
    "quality_score": 0.94,
    "domain": "general_society",
    "type": "paragraph",
    "split": "train",
    "char_length": 216,
    "word_count": 29
  }
}

Loading the dataset

Install the Hugging Face Datasets library:

pip install datasets

Load the corpus:

from datasets import load_dataset

dataset = load_dataset("ai4ug/luganda_prose")
print(dataset)
print(dataset["train"][0])

Access the individual splits:

train_dataset = dataset["train"]
validation_dataset = dataset["validation"]
test_dataset = dataset["test"]

Retrieve a text sample and its metadata:

example = train_dataset[0]
text = example["text"]
domain = example["domain"]
content_type = example["type"]
print(text)
print(domain)
print(content_type)

Data quality and limitations

We applied Unicode NFC normalization, HTML/boilerplate removal, PII scrubbing, Luganda language identification (LID) scoring, global deduplication, and split-level overlap checks across all records in this corpus.

When training or evaluating models on this corpus, please consider:

Domain and content-type distribution: Sentence-level records in grammar_and_news form the largest share of the corpus, complemented by over 134,000 multi-sentence paragraphs and news articles across society, agriculture, governance, health, business, education, and culture.
Linguistic variation: Luganda usage varies across regions, generations, and registers.
Downstream validation: Models trained on this corpus for consequential domains (such as healthcare, law, finance, or agricultural advisory) should be validated with appropriate human oversight.

Citation and attribution

If you use this dataset in research, publications, model development, or derived work, please cite it as follows:

Recommended citation

Kambale, W. (2026). Luganda Prose, Sentences and Vocabulary Corpus (Version 1.0.0) [Dataset]. AI for Uganda Initiative. Hugging Face. https://huggingface.co/datasets/ai4ug/luganda_prose

BibTeX:

@misc{kambale2026lugandaprose,
  author = {Kambale, Wesley},
  title = {{Luganda Prose, Sentences and Vocabulary Corpus}},
  year = {2026},
  publisher = {AI for Uganda Initiative},
  howpublished = {\url{https://huggingface.co/datasets/ai4ug/luganda_prose}},
  note = {Version 1.0.0. Monolingual Luganda corpus for language modelling, developed for the AI for Uganda Initiative}
}

Acknowledgements

We developed and curated this corpus at the AI for Uganda Initiative to advance African language technology and support Pearl Chat, a native Luganda conversational language model.

Author: Wesley Kambale
Initiative: AI for Uganda Initiative (https://aiforuganda.org)
Dataset: https://huggingface.co/datasets/ai4ug/luganda_prose
Project: Pearl Chat (https://pearl.aiforuganda.org)

Downloads last month: 29

Total file size:

164 MB

Models trained or fine-tuned on ai4ug/luganda_prose

ai4ug/pearl-chat-500m-it

Text Generation • 0.5B • Updated 2 days ago • 5