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 full AƔBALU-Text v1 corpus and speaker-disjoint Common Voice Kabyle splits. Every sentence is broken into lowercased ASR tokens with two parallel label sequences: which punctuation mark follows each word, and how the word is capitalised. It is the training and evaluation corpus for agbalu/Belaid-31M, which restores punctuation at 0.793 macro-F1 on the held-out test split.

It is, as far as we can establish, the first labelled punctuation restoration corpus for Kabyle or any Berber language.

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	Long-form prose (HCA), held out of training entirely

Schema

field	type	description
words	list[str]	Lowercased, unpunctuated word tokens — the form an ASR system emits
punctuation	list[str]	Per-word punctuation label: NONE, COMMA, PERIOD, QUESTION, or COLON
case	list[str]	Per-word casing label: LOWER or UPPER_INIT
source	str	Provenance tag: which corpus this sentence came from

Every list is the same length (one entry per word). Labels align to words, not subwords: words[i] takes mark punctuation[i] and initial casing case[i].

Label Scheme

Punctuation

label	character	support in test
NONE	—	20,904
PERIOD	.	4,357
QUESTION	?	861
COMMA	,	804
COLON	:	43

! and ; fold into PERIOD. Together they account for 0.41% of tokens, and the corpus contains identical sentences under both marks — the distinction is not recoverable from text alone, and collapsing them costs less than carrying a class that cannot be learned.

Casing

label	meaning
LOWER	Word begins lowercase
UPPER_INIT	First character is uppercase (sentence-initial or proper noun)

ALL_CAPS was trialled as a third class. It occurs once in dev and three times in test — too few to learn without spurious generalisation. The model that carried it emitted fourteen all-capitals words for the three real ones. Restoring an acronym to full capitals is out of scope, and the class is absent.

Decontamination

58.2% of Common Voice Kabyle transcripts also appear in AƔBALU-Text v1, because the majority are Tatoeba sentences and Tatoeba is in the corpus. Every clip whose text is found in the text corpus is removed from dev and test. This simultaneously decontaminates the Masinissa-31M encoder used by Belaid-31M, which was pretrained on the same file. The rows in the Belaid-31M evaluation are unseen by both the punctuation heads and the backbone.

A transcript without a sentence-final mark is dropped rather than labelled NONE. Those transcripts carry neither punctuation nor capitals — transcriber habit, not Kabyle grammar.

opus.nllb-kab (871,663 records) is excluded from training entirely. Its record boundaries were chosen by a bitext miner rather than a writer; a record can end mid-sentence, and sentence-final punctuation is the predicted label.

Out-of-Domain Split (ood)

The ood split is hf.imsidag.kabyle-corpus-hca — 45,028 records of continuous long-form Kabyle prose, held out of training entirely. Common Voice records are one sentence each; HCA records average 22 words across several sentences.

The Belaid-31M card reports this split explicitly alongside test. A model evaluated only in its training domain has not been evaluated.

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 (miner-cut boundaries)	—

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

Reproduction

make punctuation TASK=corpus

Reads AƔBALU-Text v1 and the Common Voice speech splits, applies decontamination, and writes the four files to data/processed/punctuation/.

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 the label annotations and the dataset release. The underlying sentences retain their upstream licences:

• Common Voice train/dev/test clips: CC0-1.0 (Mozilla Foundation).
• Text corpus sentences: mixed — see the composition table above. Every share-alike source (opus.nllb-kab, ODbL-licensed rows) was excluded before publication. No row in any published split imposes a redistribution obligation on downstream users.
• The label sequences (punctuation, case) are this project's annotation and are Apache-2.0.

Part of AƔBALU, a Kabyle corpus and model collection.