bm-text-normalization-benchmark

A small human-annotated evaluation set for Bambara (Bamanankan) orthographic normalisation: 96 real-world Bambara strings, each paired with a hand-written standard-orthography rewrite. It is the cleaned export of the finished annotations from djelia/text-normalization-benchmark.

Load

from datasets import load_dataset

# the current, whitespace-clean evaluation set
bench = load_dataset("djelia/bm-text-normalization-benchmark", "level-2", split="test")

# score one source pool at a time
asr_only = bench.filter(lambda row: row["source_dataset"] == "bambara-asr-v2")

Configs

Config	Split	Rows
level-2	test	96
default	train	96
level-1	test	81

Prefer level-2. default holds the same 96 rows but 33 of its targets carry a trailing newline; level-1 is an earlier 81-row variant.

Fields

Field	Description
source_dataset	Source pool: transcription.txt (39), Denube-final (24), bambara-asr-v2 (19), kunkado (14)
source_text	Raw Bambara string, 3-243 characters (median 43)
target_text	Human-written standard-orthography rewrite

Example: iyere lafia sa thie -> I yɛrɛ lafiya sa, cɛ́.

Notes

The task is broader than diacritic restoration: it covers punctuation, capitalisation, word re-segmentation, French code-switched material, and lexical correction where the input was garbled.

20 of the 96 rows are identity pairs, so a model that rewrites everything is penalised on a fifth of the set.

At 96 items one row is roughly one point of exact-match accuracy. Report a character-level metric (CER or normalised edit distance) alongside exact match.