KabPunct

A punctuation and capitalisation restoration corpus for Kabyle (Taqbaylit, kab, Latin script), from the AƔBALU project.

1,318,707 word-labelled sentences drawn from the AƔBALU-Text v1 corpus and speaker-disjoint Common Voice Kabyle splits. Every sentence is tokenized into lowercased ASR words with two parallel label sequences: the trailing punctuation mark following each word, and the word's capitalization state. It serves as the training and evaluation corpus for agbalu/Belaid-31M (0.793 macro-F1 on the held-out test split).

from datasets import load_dataset

ds = load_dataset("agbalu/KabPunct")

row = ds["train"][0]
# {
#   "words":       ["tecfiḍ", "fell-i"],
#   "punctuation": ["NONE",   "QUESTION"],
#   "case":        ["UPPER_INIT", "LOWER"],
#   "source":      "hf.abdelhaqueidali.kab-latn-tfng"
# }

def restore(row):
    mark_map = {"NONE": "", "COMMA": ",", "PERIOD": ".", "QUESTION": "?", "COLON": ":"}
    parts = []
    for word, punct, case in zip(row["words"], row["punctuation"], row["case"]):
        w = word[:1].upper() + word[1:] if case == "UPPER_INIT" else word
        parts.append(w + mark_map[punct])
    return " ".join(parts)

print(restore(row))
# "Tecfiḍ fell-i?"

Splits

Split	Config	Rows	Words	Description
train	default	1,262,922	12,101,835	Text corpus + Common Voice train
dev	default	5,597	29,280	Common Voice dev, decontaminated
test	default	5,160	26,969	Common Voice test, decontaminated
ood	ood	45,028	1,001,370	Continuous prose (HCA), held out of training

Schema

Field	Type	Description
words	list[str]	Lowercased, unpunctuated word tokens (raw ASR format)
punctuation	list[str]	Per-word trailing punctuation mark: NONE, COMMA, PERIOD, QUESTION, COLON
case	list[str]	Per-word casing label: LOWER or UPPER_INIT
source	str	Provenance source identifier

Labels align to whitespace-delimited words: words[i] takes trailing mark punctuation[i] and capitalization case[i].

Label Scheme

Punctuation

Label	Mark	Support in test
NONE	—	20,904
PERIOD	.	4,357
QUESTION	?	861
COMMA	,	804
COLON	:	43

! and ; fold into PERIOD during training, as together they account for 0.41% of tokens and represent terminal pauses.

Casing

Label	Meaning
LOWER	Word begins lowercase
UPPER_INIT	Initial character is uppercase (sentence-initial or proper noun)

ALL_CAPS is excluded from the single-utterance configuration because acronyms occur too infrequently in isolated Common Voice sentences (1 instance in dev, 3 in test) for robust sentence-level supervision. Full-discourse acronym restoration is addressed in agbalu/Belaid-75M.

Usage

from datasets import load_dataset

train = load_dataset("agbalu/KabPunct", split="train")
print(train[0])
# {'words': ['tecfiḍ', 'fell-i'],
#  'punctuation': ['NONE', 'QUESTION'],
#  'case': ['UPPER_INIT', 'LOWER'],
#  'source': 'hf.abdelhaqueidali.kab-latn-tfng'}

ood = load_dataset("agbalu/KabPunct", "ood", split="ood")

Surface text reconstruction helper:

MARK = {"NONE": "", "COMMA": ",", "PERIOD": ".", "QUESTION": "?", "COLON": ":"}

def render(row):
    return " ".join(
        (word.capitalize() if case == "UPPER_INIT" else word) + MARK[mark]
        for word, mark, case in zip(row["words"], row["punctuation"], row["case"])
    )

print(render(train[0]))  # 'Tecfiḍ fell-i?'

Evaluating on both test and ood provides a comprehensive assessment: test consists of single-utterance Common Voice sentences, while ood contains continuous multi-sentence literary prose averaging 1.75 sentence-final marks per record.

Decontamination

58.2% of Common Voice Kabyle transcripts appear in the general web crawl via Tatoeba. All utterances sharing strings with the training corpus were removed from dev and test splits.

Transcripts lacking sentence-final punctuation were removed from the corpus to prevent unpunctuated transcription artifacts from biasing label statistics.

opus.nllb-kab (871,663 records) was excluded from training entirely because automated bitext-mined segment boundaries frequently truncate sentences mid-clause.

Out-of-Domain Split (ood)

The ood split comprises 45,028 continuous prose paragraphs from hf.imsidag.kabyle-corpus-hca, held out completely from training. While Common Voice utterances contain a single sentence per record, HCA records average 22 words across multiple clauses and sentences.

The Belaid-31M evaluation reports metrics across both test (in-domain single utterances) and ood (out-of-domain continuous prose).

Composition

Source ID	Type	Train Rows
hf.abdelhaqueidali.kab-latn-tfng	Text corpus	647,039
hf.imsidag.kabyle-corpus-ummto	Text corpus	136,799
hf.imsidag.kabyle-raw-text	Text corpus	121,338
hf.imsidag.kabyle-corpus-ubouira	Text corpus	76,336
hf.boffire.kab-en-toponyms-sentences	Text corpus	30,584
speech.train	Common Voice 22.0 train	151,100
Others (28 sources)	Text corpus	99,726
opus.nllb-kab	Excluded (automated bitext boundaries)	—

The ood split contains hf.imsidag.kabyle-corpus-hca exclusively (45,028 rows).

Citation

@misc{agbalu_kabpunct,
  title  = {KabPunct: a Kabyle punctuation and casing restoration corpus},
  author = {AƔBALU},
  year   = {2026},
  url    = {https://huggingface.co/datasets/agbalu/KabPunct}
}

Licence

Apache-2.0 for dataset annotations and release format. Underlying text retains source licensing:

• Common Voice clips: CC0-1.0 (Mozilla Foundation).
• Text corpus sentences: Mixed permissive sources (all share-alike and ODbL sources excluded prior to publication).
• Label sequences (punctuation, case): Apache-2.0.