koussedia/nko-ocr-dataset

N'Ko OCR Dataset

Training dataset for N'Ko (ߒߞߏ) optical character recognition — the script of Manding languages (Bambara, Maninka, Dioula), ~50 million speakers in West Africa.

Built from scratch (Oct 2026) because no usable N'Ko OCR dataset existed anywhere.

Contents

• 4,422 PNG images + transcriptions (.gt.txt) + Tesseract box files (.box)
• 2,211 unique N'Ko lines (Bambara, Mali), each rendered in 2 variants (clean 40pt / noisy 32pt)
• 2 fonts: Noto Sans NKo, Afronik N'Ko
• 500 targeted lines rich in diacritics (4,677 marks) and N'Ko digits (1,504)
• corpus-nko.txt — the full text corpus (one line per file)
• LISEZ-MOI.txt — notes (French)

All images are printed text (synthetically rendered with PIL + libraqm, RTL-correct). No handwriting.

Provenance

• Authentic lines: N'Ko learning materials, ebook extraction, real N'Ko Facebook posts
• 1,221 lines recombined from authentic pairs (data augmentation)
• Text normalized to NFC (as Tesseract's tesstrain expects)

Proven results

Models trained on this dataset (koussedia/nko-ocr):

Model	CER	WER
v1	15.84%	34.84%
v2	6.26%	17.85%

License

GPLv3 — same as the models. Author: Fousseyni Diarra, Bamako, Mali.

Citation

@dataset{diarra2026nkoocrdataset,
  author = {Fousseyni Diarra},
  title  = {N'Ko OCR Dataset},
  year   = {2026},
  url    = {https://huggingface.co/datasets/koussedia/nko-ocr-dataset}
}