agbalu/KabLiterary

KabLiterary

A 28,907-paragraph, 657,713-word corpus of classical world literature in canonical Kabyle Latin orthography, from the AƔBALU project.

Eleven works — translated or originally composed in Kabyle — spanning epic poetry, gothic fiction, philosophical prose, folktales, military strategy, and 19th-century novels. The corpus is designed for language modelling, vocabulary probing, fine-tuning, and literary translation benchmarking where long-form, high-register Kabyle text is required.

from datasets import load_dataset

ds = load_dataset("agbalu/KabLiterary")
# DatasetDict({
#     'train': Dataset(27619),
#     'dev': Dataset(352),
#     'test': Dataset(936)
# })

Splits

Split boundaries are at the book level. No paragraph from a dev or test book appears in train.

Corpus Inventory

Isefra n Si Mohand u Mhand is the single original Kabyle work — oral poetry transcribed by Mouloud Mammeri and not derived from any source language. All other works are machine-translated from the source editions listed above.

Schema

Every record shares a fixed rectangular schema:

source_text is populated only for works where line-level alignment was preserved during translation — The Art of War, The Prophet, and Isefra n Si Mohand u Mhand. For the longer prose novels, paragraphs were batched during translation and source_text is empty.

Usage

Default splits

from datasets import load_dataset

dataset = load_dataset("agbalu/KabLiterary")
print(dataset["test"][4]["text"])
# 'Tella Alice tebda teɛya aṭas seg akken teqqim ɣer tama n uletma-s...'

Filtering by work or genre

from datasets import load_dataset

train = load_dataset("agbalu/KabLiterary", split="train")

# French 19th-century novels only
french_novels = train.filter(
    lambda row: row["book_id"] in ("le-comte-de-monte-cristo", "les-miserables")
)
print(f"{len(french_novels):,} paragraphs, {sum(french_novels['word_count']):,} words")

# Records with 1-to-1 source alignment
parallel = train.filter(lambda row: len(row["source_text"]) > 0)
print(f"{len(parallel):,} line-level parallel records")

Full plain-text books

All 11 works are included in the repository's books/ directory:

git clone https://huggingface.co/datasets/agbalu/KabLiterary
ls KabLiterary/books/
cat KabLiterary/books/the-prophet.txt

Curation

Translation

The ten translated works were produced by NLLB-based machine translation and subjected to a multi-stage quality pass before inclusion:

1. 1.Repetition pruning. 276+ beam-search generation loops were detected and removed across all ten texts. No tail-degenerate paragraph survives in any split.
2. 2.Heading localisation. Residual French chapter headings in Les Misérables and other works were translated into standard Kabyle Latin (e.g. L'évêque travaille → Aselway ixeddem, L'Alouette → Tiwiziwt).
3. 3.Source-language artifact removal. Any paragraph retaining a title or quotation in English or French was re-translated or manually corrected to Kabyle.

Orthography

All text is normalised to AƔBALU canonical Kabyle Latin orthography (ɣ ɛ ḥ ḍ ṣ ṭ ẓ ṛ č ǧ). A census across all 28,907 records confirmed:

• —zero Greek epsilon homoglyphs (ε U+03B5)
• —zero soft hyphens (\u00ad)
• —NFC normalisation throughout

HTML entities (" and variants) were repaired to standard punctuation during the quality pass.

Known Limitations

• —Machine-translated syntax. Ten of eleven works are machine-translated. The text is fluent and orthographically correct, but phrase-level structure carries influence from the French or English source. No human post-editing was performed beyond the artifact and heading corrections described above.
• —No 1-to-1 paragraph alignment in prose novels. Longer works (Monte-Cristo, Les Misérables, Dracula, The Arabian Nights) were translated with multi-sentence batching. Paragraph boundaries do not map to individual source sentences.
• —Archaic and specialised vocabulary. Works such as Confessions of Saint Augustine and The Odyssey require theological and classical vocabulary that has no established Kabyle equivalent. The translations use loan-adapted morphological frames; the decisions are internally consistent but not reviewed against a Kabyle lexicographic standard.
• —Single source edition per work. Each translation was produced from one source edition. Variant readings, abridgements, or multi-edition differences were not considered.

Reproduction

make release-kabliterary

The dataset is assembled by tools.build_kabliterary from the verified texts in artifacts/translations/ and data/documents/.

Citation

@misc{agbalu_kabliterary,
  title  = {KabLiterary: A 657K-Word Corpus of Classical Literary Translations in Kabyle},
  author = {AƔBALU},
  year   = {2026},
  url    = {https://huggingface.co/datasets/agbalu/KabLiterary}
}

Licence

CC-BY-SA-4.0. Source texts are public domain (Project Gutenberg and digitised archives). Translations and the assembled dataset are produced by the AƔBALU project under CC-BY-SA-4.0. Attribution is required; derivative datasets must carry the same licence.

Part of AƔBALU, a Kabyle NLP corpus and model suite.