text-normalization-benchmark

The raw Argilla 2.8.0 export of a Bambara (Bamanankan) text-normalization project: 160 records from four in-house corpora, each with the annotator's standard-orthography rewrite. 96 carry a submitted response; 64 were discarded. For a ready-to-score evaluation set, use djelia/bm-text-normalization-benchmark, the cleaned export of the 96 finished annotations.

The repo is gated: request access on the Hub and run hf auth login.

Load

from datasets import load_dataset

ds = load_dataset("djelia/text-normalization-benchmark", split="train")  # 160 rows

gold = ds.filter(lambda r: r["normalized_text.responses.status"][0] == "submitted")
pairs = [(r["source_text"], r["normalized_text.responses"][0]) for r in gold]  # 96

The .argilla/ folder carries the schema, so the project reopens with rg.Dataset.from_hub(...).

Fields

Column	Description
source_text	The raw Bambara sentence to normalize
normalized_text.responses	List with the annotator's normalized text
normalized_text.responses.status	submitted (96) or discarded (64)
source	In-house corpus: bambara-asr-v2 50, Denube-final 50, transcription.txt 40, kunkado 20
status	completed (96) or pending (64)

Plus source_index, record identifiers, timestamps, the annotator UUID, and normalized_text.suggestion — a UI pre-fill equal to source_text in every row.

Notes

Filter on the response status before treating anything as gold: a discarded response still carries a value, in 62 of the 64 cases a copy of source_text.

33 of the submitted targets end in a trailing newline that no source_text contains; strip it before any string-match metric.

Some rows carry transcription markers such as <INCOMPRÉHENSIBLE>.