Maloba

Documents corpus · Reference version

dataset-qwen-vl-lingala-qlora-vf

Multimodal training set: 417 training examples, 50 test examples (text-region images and transcriptions).

View on Hugging Face (opens in a new window)All corpora

How it was built

  • Images of Lingala text lines taken from the SIL Congo collection (printed documents and manuscripts), annotated as regions then transcribed.
  • 417 training examples (278 lines) and 50 test examples; the tokens ɔ ɛ Ɔ Ɛ were added to the model.
  • 1,323 unique validated transcriptions and 1,282 harmonised lines, kept outside the published training set.

The corpus, as charts

Corpus sizes comparedThe project's four reference corpora, on the same scale. The speech corpus dominates in number of items; translation, the smallest, is the one that moved its model the most.
05 06710 13415 20120 268Speech — segments18 767Text — annotated examples4 818Documents — examples467Translation — human pairs399

Source : Livrable 4, chiffres arrêtés au 2026-09-30

Show the chart data
CorpusItems
Speech — segments18 767
Text — annotated examples4 818
Documents — examples467
Translation — human pairs399

Limits

  • Small test set (50 examples); uneven image quality.

Citation

Congo Digital Services (2026). dataset-qwen-vl-lingala-qlora-vf. Hugging Face.