Loading datasets...
Release Date: 10/12/2026
Format: JSONL
Size: 338.41 KB
MALOBA French–Lingala translation pairs, configuration h2, published by Congo Digital Services (CDS SARL) for the MALOBA project, in partnership with UNDP Republic of Congo and the Ministry of Posts, Telecommunications and the Digital Economy. This is a text translation dataset, not an audio dataset. It contains 4,874 French → Lingala sentence pairs. The Lingala side follows the harmonised MALOBA spelling convention. The pairs have three origins, given in the column "origine": - 399 human translations (TRAD set); - 432 questions translated manually by annotators; - 4,043 questions pre-filled by a machine translation model, then validated or corrected by annotators (post-editing). The dataset is split into 4,763 training, 32 validation and 79 test pairs. Validation and test contain only human translations from the TRAD set; all question pairs are in training. The split is made by French sentence: all Lingala versions of a French sentence stay on the same side, and no training sentence appears in validation or test, in exact or near-duplicate form. A French question can have up to four Lingala versions (1,348 distinct French questions in training). 104 identical pairs were removed. The Lingala spelling was harmonised by an automatic and deterministic pipeline. No text was rewritten by a person or a generative model during harmonisation. The original Lingala is kept in "output_brut" and each change is listed in "modifications". Known limitations: the model used for pre-annotation has not been identified; the Lingala translations have not all been reviewed by linguists.
Licensing
Nwulite Obodo Open Data Licence 1.0 (NOODL-1.0)
Restrictions/Special Constraints
Open to all users who accept the NOODL-1.0 licence. Access requests are approved automatically so that every user is identified and accepts the licence before downloading. Users based outside Africa and developing countries must contact Congo Digital Services and agree terms before any use, as NOODL-1.0 requires. Attribution to Congo Digital Services and to the MALOBA project is required. The 79 test rows (dans_test_v1 = true) must never be used for training.
Forbidden Usage
Training on the 79 test rows (dans_test_v1 = true). Any use that does not comply with NOODL-1.0, including use outside Africa and developing countries without prior agreement with Congo Digital Services.
Ethical Review
Nature of the data. Text only: French sentences and their Lingala translations. No audio, no voice recordings and no personal data about speakers. Project and source material. The dataset was produced within MALOBA, the national project for the digitisation of Lingala, carried out by the consortium WYMEE / Congo Digital Services / SACLI under UNDP contract no. 10479844. The French source sentences and contexts were written by the MALOBA team itself; no third-party material was used. Contributors and compensation. The translations were produced in Label Studio by Lingala-speaking annotators following a written annotation guide. Annotators were either CDS employees, paid a salary, or external contributors engaged by CDS and paid per task or through allowances, in particular for linguistic validation. They took part as paid annotators, knowing the data would be used to build Lingala language technology. They were trained in the annotation protocol on the job; an AI training programme in three sessions with a certificate is planned. Annotator identities are not published. Risks and safeguards. The main risk is quality: part of the pairs are machine translations post-edited by annotators. The column origine documents this, so users can keep human translations only. Anyone who finds content that should be removed can write to [email protected].
Intended Use
Training and evaluation of French → Lingala machine translation.
TYPE OF DATASET Text dataset for French → Lingala machine translation: sentence pairs only, no audio.
COLLECTION AND PROCESSING The pairs were produced in Label Studio by Lingala-speaking annotators of the MALOBA team, following a written annotation guide. Some questions were translated manually; others were pre-filled by a machine translation model, then validated or corrected by an annotator. The origin of each pair is given in the column origine, so users can filter by origin.
SIZE AND FORMAT 4,874 pairs in JSONL: 4,763 train, 32 validation, 79 test. 1,636 distinct French sentences in train.
VERSIONS h1: the 399 human translations (TRAD), split 288 / 32 / 79. h2 (this version): h1 plus 4,475 question pairs in train; validation and test are unchanged.
ACCESS The dataset is open to anyone who accepts NOODL-1.0, not only to CDS members. Access requests are approved automatically: this records who downloads the data and ensures every user accepts the licence, including attribution to Congo Digital Services and to the MALOBA project. Users outside Africa and developing countries must contact Congo Digital Services before any use. Access can be revoked if the licence is not respected.
