monilingual-texts

Plain single-column text corpora in Bambara (Bamanankan), French and English, laid out with one split per language rather than train/validation/test. Intended for tokenizer training, language modelling, continued pretraining and vocabulary statistics.

Load

from datasets import load_dataset

bambara = load_dataset("djelia/monilingual-texts", "default", split="bm")
french  = load_dataset("djelia/monilingual-texts", "default", split="fr")
english = load_dataset("djelia/monilingual-texts", "default", split="en")

# the tone-marked lexicon slice, row-aligned across its two splits
lex_bm = load_dataset("djelia/monilingual-texts", "bamadaba", split="bm")
lex_fr = load_dataset("djelia/monilingual-texts", "bamadaba", split="fr")
pairs  = list(zip(lex_bm["text"], lex_fr["text"]))

Configs

Config	Split	Rows	Mean chars
default	bm	98,062	77.9
default	fr	84,381	92.8
default	en	13,704	73.7
bamadaba	bm	4,854	27.3
bamadaba	fr	4,854	32.0

One field, text.

Notes

The splits are languages, not data partitions. load_dataset without a split argument returns a DatasetDict keyed by language code, and there is no held-out portion.

The default splits are not aligned with each other and row order is shuffled, so zipping them does not recover sentence pairs. bamadaba is the exception: its two splits preserve row order, so zipping them gives valid Bambara-French phrase pairs. Its content is short dictionary-style material carrying tone diacritics, which the rest of the corpus mostly lacks.

Roughly 30% of default/bm and 27% of default/fr are exact duplicates of another row. Deduplicate before any frequency count.