We developed and curated ai4ug/luganda_qa_pairs as a Luganda-language dataset to support research and development in question answering, conversational language modelling, and supervised fine-tuning (SFT). It contains 68,421 question and answer pairs comprising 8,898,295 words (including 3,860 multi-turn 5-message dialogues), covering a broad range of subjects across education, science, agriculture, healthcare, business, language, history, culture, and everyday life.

We built this dataset at the AI for Uganda Initiative as part of our work on Pearl Chat, a native Luganda conversational language model. Our goal is to provide clean, structured examples for training and evaluating language models in Luganda, a language that remains underrepresented in large-scale natural language processing research.

Each record includes a question, its corresponding answer, a combined text representation, and a single-turn or multi-turn conversational representation (system, user, and assistant). We also included fields for task classification, domain labels, language metadata, text statistics, quality scores, and dataset split assignments.

Dataset at a glance

Property	Description
Total records	68,421
Total words	8,898,295
Multi-turn (5-message) dialogues	3,860
Language	Luganda
ISO 639-3 code	`lug`
Script	Latin (`Latn`)
Text normalization	Unicode NFC
Primary tasks	Question answering, conversational modelling, instruction tuning
Record structure	Question, answer, text, conversations, metadata
Splits	Train, validation, test
Split proportions	80%, 10%, 10%
Primary project	Pearl Chat
Author & Maintainer	Wesley Kambale (AI for Uganda Initiative)
License	Apache 2.0

Intended uses

We designed this dataset to support the following applications:

Supervised fine-tuning: Training language models to respond accurately to instructions and questions in Luganda.
Conversational modelling: Adapting autoregressive language models to natural single-turn and multi-turn Luganda dialogue.
Question answering: Training systems to generate grounded responses to Luganda-language questions.
Language research: Studying Luganda vocabulary, syntax, text generation, and conversational behaviour.
Evaluation: Measuring model performance on held-out Luganda question and answer pairs.
African language technology: Advancing low-resource NLP and locally relevant language applications.

You can use this dataset independently or alongside other appropriately licensed Luganda corpora.

Dataset composition

Domain distribution

We categorised all records by subject domain across Ugandan society and daily life:

Domain	Records	Share
Education and literacy	12,447	18.2%
Science and technology	10,932	16.0%
Agriculture	10,420	15.2%
Health and medicine	9,578	14.0%
Business and finance	7,970	11.6%
Grammar and linguistics	4,492	6.6%
History, culture and kingdoms	3,048	4.5%
Geography and tourism	2,967	4.3%
Civic life and law	1,915	2.8%
Family, parenting and youth	1,556	2.3%
Sports and entertainment	1,148	1.7%
Urban life and transport	869	1.3%
Encyclopedic knowledge	431	0.6%
Science and education	299	0.4%
Society and news	183	0.3%
Culture and society	161	0.2%
Society and occupations	2	<0.1%
Society and civics	1	<0.1%
Vocabulary and grammar	1	<0.1%
Society and finance	1	<0.1%
Total	68,421	100%

Task modes

Each record includes a mode field indicating its intended task:

Mode	Records	Description
`dialog`	63,016	Single-turn and multi-turn conversational question answering and response generation
`continuation`	4,928	Text, literary, story, and sentence continuation
`translate`	477	Meaning, proverb (`engero`), and lexical translation/explanation tasks
Total	68,421

Dataset splits and deduplication

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

Split	Records	Share	Words
Train	54,731	80%	7,122,060
Validation	6,839	10%	891,350
Test	6,851	10%	884,885
Total	68,421	100%	8,898,295

Split construction

To prevent train-test contamination and memorization, we used Union-Find connected-component clustering to group related question and answer pairs before assigning records to splits.

We linked pairs whenever they shared either a canonicalised answer or a normalised question prefix, and assigned connected records to the same split so closely related examples do not appear across both training and evaluation sets.

Our verification checks confirm:

Question overlap across the three splits: 0.0%.
Answer overlap across the three splits: 0.0%.
Exact cross-dataset overlap with ai4ug/luganda_prose: 0.0%.

For reproducible experiments, please preserve the supplied split assignments and avoid moving evaluation examples into training data.

Data format and schema

Each record contains the following principal fields:

Field	Type	Description
`id`	String	Unique record identifier
`mode`	String	Task category (`dialog`, `continuation`, `translate`)
`domain`	String	Subject-domain classification
`split`	String	Dataset partition (`train`, `validation`, `test`)
`question`	String	Luganda-language question or prompt
`answer`	String	Corresponding response
`text`	String	Combined question and answer text
`added`	String	Record date in `YYYY-MM-DD` format
`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
`conversations`	Array	Role-labelled conversational messages
`metadata`	Object	Structured record-level metadata

Example record

{
  "id": "lug_qa_000001",
  "mode": "dialog",
  "domain": "agriculture",
  "split": "train",
  "question": "Nga nnamaze okukungula muwogo gwange...",
  "answer": "Kituufu nti ebitooto by'omuwogo...",
  "text": "Ekibuuzo: Nga nnamaze okukungula muwogo gwange...\nOkuddamu: Kituufu nti ebitooto by'omuwogo...",
  "added": "2026-10-15",
  "language": "lug",
  "script": "Latn",
  "quality_score": 0.99,
  "char_length": 482,
  "word_count": 64,
  "conversations": [
    {
      "role": "system",
      "content": "Oli Pearl Chat, omuyambi wa AI ayogera Oluganda olutuufu era annyonnyola ebintu mu bujjuvu."
    },
    {
      "role": "user",
      "content": "Nga nnamaze okukungula muwogo gwange..."
    },
    {
      "role": "assistant",
      "content": "Kituufu nti ebitooto by'omuwogo..."
    }
  ],
  "metadata": {
    "language": "lug",
    "script": "Latn",
    "quality_score": 0.99,
    "domain": "agriculture",
    "mode": "dialog",
    "split": "train",
    "char_length": 482,
    "word_count": 64
  }
}

Conversational representation

The conversations field provides a sequence of role-labelled messages:

system: Instructions defining the assistant's persona and task behaviour in Luganda.
user: The question or instruction in Luganda.
assistant: The corresponding response in Luganda.

You can use conversations directly in conversational fine-tuning pipelines or use the top-level question, answer, and text fields for standard question-answering and text-generation workflows.

Loading the dataset

Install the Hugging Face Datasets library:

pip install datasets

Load the dataset:

from datasets import load_dataset

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

Access a particular split:

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

Access the question and answer fields:

example = train_dataset[0]
question = example["question"]
answer = example["answer"]
print("Question:", question)
print("Answer:", answer)

Access the conversational representation directly:

example = train_dataset[0]
for message in example["conversations"]:
    print(message["role"], message["content"])

Data quality and limitations

We applied Unicode NFC normalization, HTML/boilerplate removal, PII scrubbing, Luganda language identification (LID) quality scoring, exact deduplication, and split-level overlap checks across every record in this dataset.

When training or deploying models on this dataset, please keep the following considerations in mind:

Factual accuracy: For consequential domains such as healthcare, law, finance, and agricultural extension, model outputs should be reviewed and validated by qualified professionals.
Linguistic variation: Luganda usage varies across speakers, regions, contexts, and registers.
Domain balance: Education, science and technology, agriculture, healthcare, and business represent the largest share of records.

Citation and attribution

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

Recommended citation

Kambale, W. (2026). Luganda Question and Answer Pairs (Version 1.0.0) [Dataset]. AI for Uganda Initiative. Hugging Face. https://huggingface.co/datasets/ai4ug/luganda_qa_pairs

BibTeX:

@misc{kambale2026lugandaqa,
  author = {Kambale, Wesley},
  title = {{Luganda Question and Answer Pairs}},
  year = {2026},
  publisher = {AI for Uganda Initiative},
  howpublished = {\url{https://huggingface.co/datasets/ai4ug/luganda_qa_pairs}},
  note = {Version 1.0.0. Dataset for Luganda-language question answering and conversational modelling, developed for the AI for Uganda Initiative}
}

Acknowledgements

We developed and curated this dataset at the AI for Uganda Initiative to advance African language technology and power 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_qa_pairs
Project: Pearl Chat (https://pearl.aiforuganda.org)

Downloads last month: 28

Total file size:

110 MB

Models trained or fine-tuned on ai4ug/luganda_qa_pairs

ai4ug/pearl-chat-500m-it

Text Generation • 0.5B • Updated about 8 hours ago