Resources
Guides, code, deliverables
Everything needed to reuse, reproduce or extend the work: spelling conventions, tools, public repositories, reports.
Guides and tools
Segmentation, word-by-word transcription, ɔ and ɛ, tones, French words, tags. The annotator training document. It is not released on this site.
A deterministic, dependency-free program that reproduces the harmonised corpus columns exactly; majority lexicon of 11,382 entries. It is published with corpus V3.2, and has no standalone repository.
More than 1,400 automated tests; replays every measurement from the published weights (reading C, bootstrap by recording, FLORES reference). The repository remains private.
Public repositories
Only public repositories are listed. GitHub organisation: Congo-Digital-Services (opens in a new window); Hugging Face organisation: Congo-digital-service (opens in a new window).
Conversation interface and integrations of the Maloba models.
View the repository (opens in a new window)LibreChat (dérivé CDS, interface LoBAI)The CDS adaptation of LibreChat: the LoBAI interface and its Maloba agents.
View the repository (opens in a new window)maloba-ai-model-releasesModel releases: versions, release notes, loading scripts.
View the repository (opens in a new window)The models and datasets are described on the Models / Datapages, with their Hugging Face links and licences.
Reports and deliverables
Automated and human reports, with their supporting files. Annex C of Deliverable 4.
Training reports for the five models, with their logs. Annex D of Deliverable 4.
Protocol P1: sampling rules, questions put to validators, verdict counting and agreement measurement, all fixed before the campaigns.
Run logs, raw scores, training curves and pseudonymised validator responses, report by report.
Table of evaluations and training runs, each with its download link.
Corpus and processing; models and access point; capabilities and end of mission.
Press kit
Maloba and CDS logos, Voice-a-thon photographs (with permission) and a project fact sheet: on request from the Contact page. Partner logos are released only with their agreement.