MOH749/bocalantics-wolof-eval
Bocalantics Wolof eval set The held-out Wolof test rows from MOH749/Bocalantics-2.0, with audio attached, so a candidate model can be scored without re-materialising anything. 3,557 rows, 5.38 hours. Why it exists "Beats unadapted Whisper" is not a result for Wolof. Whisper has seen 2 of the parent corpus's 26 languages, so beating it is arithmetic rather than evidence. The bar is the best published model for the language -… See the full description on the dataset page: https://huggingface.co/datasets/MOH749/bocalantics-wolof-eval.
Bocalantics Wolof eval set
The held-out Wolof test rows from MOH749/Bocalantics-2.0, with audio attached, so a candidate model can be scored without re-materialising anything.
3,557 rows, 5.38 hours.
Why it exists
"Beats unadapted Whisper" is not a result for Wolof. Whisper has seen 2 of the parent corpus's 26 languages, so beating it is arithmetic rather than evidence. The bar is the best published model for the language - asr-africa/w2v2-bert-wolof-kallaama-42-hours was trained on 42 h of Kallaama, a corpus this index also holds - scored on these exact rows, with the same normaliser and the same metric.
Rows by source
Two caveats that belong on any number from this set
Grouping. Wolof's split rests on a mix of card-declared speaker identity, recording session, and - for a shrinking remainder - text alone. A WER quoted without its grouping tier is not a result; reports/splits_report.json in the parent corpus carries the tier per language.
Join rate. These are the test rows whose audio could be re-attached from upstream, not every test row. Sources differ sharply: Kallaama joins at 100% because it is ours and the transcripts match exactly, while public aggregators join lower. The set is therefore representative of what is recoverable, which is not quite the same as representative of the corpus.