tekkilab/pulaar-eval-data

INTERNAL — Pulaar ASR evaluation corpus (temporary)

Private, internal evaluation only. Do not redistribute, do not make public, delete after the evaluation. Two of the six upstream datasets declare no licence at all, which means all rights reserved — a private copy for evaluation is defensible, a public release of the same rows is not. See LICENSE_NOTICE.md.

Rebuilt from the original public sources on 2026-10-28, after the server holding the previous corpus was lost. The rebuild reproduced the previous corpus row for row (58983 rows, 183.49 h), so the pipeline is reproducible from a bare machine and six public sources with no surviving private artefact in the chain.

Contents

split	rows	hours
train	54651	173.86
dev	1317	2.79
test	3015	6.85
total	58983	183.49

463 speakers. 16 kHz mono FLAC embedded in the parquet, one utterance per row, nothing shorter than 0.4 s or longer than 30 s, no empty transcripts.

How the splits were cut

The hold-out unit is the whole broadcast for kallaama_src and the speaker wherever a speaker id exists, so no speaker and no recording appears in two splits (verified: 0 groups span a split boundary). Dev/test utterances whose exact transcript also appears in train were then purged — dev 62, test 102 — because the n-gram language model is built from train text, which makes a repeated reference easier than it looks.

Known limitation: abdouaziz_multi_dialect ships no speaker metadata, so the seven varieties it supplies cannot be speaker-disjoint. Per-variety counts are in stats.json under split_provenance.rows_without_speaker_metadata.

Both Kallaama derivatives (ngia/ASR_pulaar, Alioukun/kallaama-pulaar-filtered) are excluded: their recordings are exactly the 83 SLR151 broadcasts that kallaama_src already covers from the authoritative STM annotations, so they add no new speech, and the re-segmentation was the source of the cross-split contamination measured earlier (7.6–9.2 WER points on radio).

Scoring pulaar_asr_v4 against this test split

in_v4_train marks every row whose exact transcript was in v4's training data. v4 predates this rebuild and was cut under different split boundaries, so 32% of the test split is not unseen by v4. Filter in_v4_train == False for an honest number. A model trained on this corpus's own train split needs no such filter.

The overlap is very uneven, so the two cleanest slices are also the two largest:

variety	test rows	in v4 train
unknown	527	9	2%
pulaar	489	40	8%
pular	273	66	24%
borgu	300	109	36%
adamawa	229	103	45%
bororro	299	152	51%
caka	299	152	51%
maacina	299	168	56%
liptako	300	173	58%

Sources

source	rows	upstream	licence
abdouaziz_multi_dialect	19916	abdouaziz/fulaa_pulaar	none declared
batbaina_fub_agri	799	BATBAINA/fulfulde-dataset-asr-agri	none declared
guizme_adlam	51	guizme/adlam_fulfulde	cc-by-4.0
guizme_pulaar	130	guizme/Pulaar	cc0-1.0
kallaama_src	14441	Kallaama Pulaar SLR151	cc-by-4.0
waxal_ful_asr	23646	google/WaxalNLP	cc-by-sa-4.0

20715 of 58983 rows (35%) come from sources that declare no licence. That is why this repository is private.

Load

from datasets import load_dataset
ds = load_dataset("tekkilab/pulaar-eval-data", split="test")        # splits: train / dev / test

datasets 4.x and 5.x decode audio through torchcodec, which is not a dependency of datasets itself. Either pip install torchcodec, or skip it and read the FLAC bytes directly — which is all a row contains anyway:

import io, soundfile as sf
from datasets import load_dataset, Audio

ds = load_dataset("tekkilab/pulaar-eval-data", split="test").cast_column("audio", Audio(decode=False))
y, sr = sf.read(io.BytesIO(ds[0]["audio"]["bytes"]), dtype="float32")   # 16 kHz mono

Every field is documented in sources.json, and every filter and drop reason in stats.json.