Congo-digital-service/Dataset-Lingala-Alpaca-LLM

Lingala Alpaca LLM Dataset

This dataset contains human-generated instruction-following data in Lingala, inspired by the Alpaca dataset format, covering five task categories (formal, pedagogical, summary, urban, translation) across five public-interest domains, with a balanced representation of stylistic categories and instruction types.

General Statistics

• Total volume: 6,664 lines in total.
• Number of unique source texts: 1,772.
• Structure: The dataset is divided into 5 instruction types, with a total of 6,100 context occurrences. This differs from the final number of lines because some contexts are intentionally reused across different tasks.

Distribution by Instruction Type

The dataset is organized into five instruction categories:

• Formal (tasks_frm.json): 1,425 occurrences, covering administrative and standard responses.
• Pedagogical (tasks_pdg.json): 1,425 occurrences, covering explanation and teaching tasks.
• Summary (tasks_rsm.json): 1,425 occurrences, covering concise summarization tasks.
• Urban (tasks_urb.json): 1,425 occurrences, covering everyday and conversational language.
• Translation (tasks_trd.json): 400 occurrences, covering translation of source texts.

Distribution by Domain and Complexity

Domains Covered

The dataset covers the following domains:

• Agriculture
• Governance / Administration
• Health
• Informal Sector / Urban
• Education Each domain contains 1,220 context occurrences.

Complexity Levels

• Basic: 2,586 contexts (42.4%).
• Intermediate: 1,952 contexts (32.0%).
• Complex: 1,562 contexts (25.6%).

Dataset Structure

Splits

The final published dataset is divided into two splits:

Split	Lines
Train	5,700
Test	964
Total	6,664

Fields

Each example contains the following fields:

• instruction
• input
• output
• source_project

Datasheet

Datasheet following the template shared by UNDP (after Gebru et al., Datasheets for Datasets, and Bender & Friedman, Data Statements for NLP).

Item	Description
Core purpose	Instruction tuning of language models in Lingala; training data of the Llama 3.1 Lingala model.
Motivation	Lingala has very few open instruction and parallel texts. Built by CDS in the MALOBA project (digitisation and acceleration of Lingala), with UNDP Republic of Congo and the Ministry of Posts, Telecommunications and the Digital Economy.
Curation rationale	Five task types and five public-interest domains, balanced by design (1,425 occurrences per task type except translation, 400).
Language / variety	Lingala (ln), with French on the source side of translation tasks. Original spelling, not harmonised.
Data subject demographics	The French source texts were written by the CDS team and institutional collaborators of the MALOBA project. No personal data about them is included.
Annotator demographics	A team of Lingala-speaking annotators, mostly graduates of the Faculté des Lettres et des Langues Vivantes, Université Marien Ngouabi (UMNG), trained in AI and IT and paid for their work by CDS. Identities and detailed sociodemographic data are not published.
Speech situation	Written text (not speech): instructions and answers in five registers (formal, pedagogical, urban, summary, translation) across agriculture, governance and administration, health, the informal sector and urban life, and education.
Composition	6,664 rows: 5,700 train, 964 test. The train split holds 3,902 distinct (instruction, input, output) rows: some rows are repeated.
Collection process	Source texts written by the CDS team and collaborators; Lingala instructions and answers produced or corrected by annotators in Label Studio.
Preprocessing / cleaning	Cross-batch deduplication of source texts (1,772 unique). Spelling not harmonised; see maloba-llm-alpaca-harmonise for the harmonised translation data.
Uses	Instruction tuning and evaluation in Lingala. Evaluate on the test split only.
Distribution	Hugging Face, gated access. Licence NOODL-1.0, full text in LICENSE.md.
Maintenance	Kept unchanged as the record of the data that trained the Llama 3.1 Lingala model. Maintained by CDS (contact@congo-digital.com).
Legal & ethical compliance	No personal data about annotators or authors is included.
Provenance	MALOBA Label Studio annotations → Alpaca format → this dataset → training of Llama-3.1-8B-Instruct-Lingala-QLoRA.

Dataset Creation

Source Data: instructions and responses written by the CDS team and institutional collaborators (UNDP, partner universities), based on 1,772 unique source texts after cross-batch deduplication.

Format and Purpose

The dataset is published in an instruction-following format compatible with the Hugging Face datasets library. The data is intended for training and fine-tuning AI language models to generate responses in Lingala adapted to different linguistic registers and task types. The dataset covers:

• Formal communication
• Pedagogical explanations
• Everyday and urban conversation
• Text summarization
• Translation It also provides coverage across several public-interest domains, including agriculture, governance and administration, health, education, and the informal/urban sector.

Considerations for Using the Data

This dataset intentionally covers 5 broad domains (agriculture, governance, health, informal sector, education) but remains modest in size (6,664 examples) for a low-resource language — users should expect limited lexical and stylistic coverage outside these domains.

Usage

To load this dataset:

from datasets import load_dataset
dataset = load_dataset("Congo-digital-service/Dataset-Lingala-Alpaca-LLM")
print(dataset)

Language

• Lingala (ln)

License

This dataset is published under the Nwulite Obodo Open Data License (NOODL-1.0), a license designed for the equitable sharing of African datasets.

This dataset, Dataset-Lingala-Alpaca-LLM (Congo-digital-service/Dataset-Lingala-Alpaca-LLM), was created by Congo Digital Services (CDS SARL) (https://www.congo-digital.com/) in August 2026 in collaboration with the MALOBA community (https://maloba.congo-digital.com/), the Faculté des Lettres et des Langues Vivantes, Université Marien Ngouabi (UMNG), the Ministry of Posts, Telecommunications and Digital Economy of the Republic of Congo, and UNDP Congo. It is licensed under the Nwulite Obodo Open Data License (https://licensingafricandatasets.com/nwulite-obodo-license).

In exchange for use of this dataset, users from developing countries are required to provide: Use of the dataset (no additional benefit required). Users from high-income countries or commercial entities are additionally required to publicly acknowledge and credit the Maloba Project (UNDP Republic of Congo — language digitalisation initiative) in any publication, product, model, or output derived from this dataset. To fulfil this requirement, contact contact@congo-digital.com or visit https://www.congo-digital.com/contact.

Creators

• Congo Digital Services (CDS SARL) — https://www.congo-digital.com/
• In collaboration with:
  • the MALOBA community — https://maloba.congo-digital.com/
  • SIL Congo
  • the Faculté des Lettres et des Langues Vivantes, Université Marien Ngouabi (UMNG)
  • the Ministry of Posts, Telecommunications and Digital Economy of the Republic of Congo
  • UNDP Congo
• Created in August 2026