bambara-mt-v2

An aggregated Bambara (Bamanankan, bm / bam_Latn) machine-translation corpus pairing Bambara with French and English, assembled from eight upstream sources.

Load

from datasets import load_dataset

# aligned table with provenance
mt = load_dataset("djelia/bambara-mt-v2", "default", split="train")

# directional training pairs
pairs = load_dataset("djelia/bambara-mt-v2", "source_target_style", split="train")

Configs

Config	Rows	Splits	Shape
default	98,086	train	one row per sentence
source_target_style	196,124	train, validation, test	one row per translation direction

Both configs hold the same corpus. Use default for the aligned table, source_target_style for directional training pairs.

Fields

default:

Field	Description
bm	Bambara sentence
en	English translation, or an empty string
fr	French translation, or an empty string
source	Upstream slice the row came from

source_target_style: source_lang, target_lang, source_text, target_text, with ISO-639-3 + script codes (bam_Latn, fra_Latn, eng_Latn).

Notes

Each default row carries exactly one of en or fr, never both; the unused column holds an empty string rather than null.

source values: bayelemabaga 46,976, lafand 27,104, guerre_des_griots 6,194, bamadaba 4,854, egafe 4,702, google_smol_gatitos 4,000, google_smol_smldoc 3,393, google_smolsent_en_bam 863.

22,341 of the 98,086 default rows are exact duplicates of another row, almost all in the lafand slice. Deduplicate before training or before computing any corpus statistic.

For evaluation, use djelia/bambara-mt-benchmark.