tatoeba-kabyle-mono-cleaned

Cleaned and quality-assessed monolingual Kabyle corpus extracted from Tatoeba.

---

Summary

This dataset contains sentences from Tatoeba tagged as Kabyle (lang == "kab"), processed through a multi-layer linguistic filtering pipeline combining orthographic normalization, language identification (GlotLID v3 + DistilBERT Kabyle/Tachelhit classifier), code-switching detection (MaskLID), and lexical validation (Kabyle Hunspell dictionary). Only the subset qualified as pure is published here as the default train split.

Language: Kabyle (kab) -- standardized Latin script (INALCO 1996 / KIM).
Source: Tatoeba sentences.csv
Data license: CC BY 2.0 FR (Tatoeba)
Code / card license: MIT
Format: HuggingFace datasets (Arrow)
Author: boffire

---

Cleaning Pipeline

1. Orthographic Normalization

• Correction of character false friends (Greek, Cyrillic, Turkish, etc.): ε → ɛ, γ → ɣ, Γ → Ɣ, Σ → Ɛ, etc.
• Removal of forbidden characters and normalization of French diacritics.
• Unicode NFC normalization and whitespace cleanup.

2. Berber Validation (DistilBERT)

• Model: boffire/distilbert-kabyle-tachelhit-classifier-v2
• Adaptive thresholds based on sentence length.
• Classification: kab (Kabyle) vs tach (Tachelhit).
• Scope: first 10,000 sentences analyzed; remainder flagged as SKIPPED.

3. Language Identification (GlotLID v3)

• Model: cis-lmu/glotlid
• Confidence threshold: >= 0.95 for the published subset.
• Allowlist fallback: predictions outside {kab, fra, ara, eng, ber, shi} are treated as noise and fall back to the sentence-level language.

4. Code-Switching Detection (MaskLID)

• Token-level analysis combining GlotLID, Kabyle Hunspell (Belkacem77), prefix stripping for verbal forms, and an auxiliary vocabulary.
• Detection of isolated borrowings and IOB2 annotation.
• CS threshold: > 15% non-Kabyle tokens.

5. "Pure" Filter

The published subset retains only sentences that simultaneously satisfy:

• is_cs == False (no code-switching)
• len(contaminants) == 0 (no false friends or forbidden characters)
• is_mismatch == False (no GlotLID=kab / DistilBERT=tach divergence)
• berber_status in {"HIGH_CONF", "SKIPPED"}
• glot_conf >= 0.95

---

Statistics

Metric	Value
Source sentences (Tatoeba lang=kab)	789,712
Retained sentences (pure subset)	756,774
Conservation rate	95.8%
Hunspell coverage (token-level)	85.1% (3,173,402 / 3,730,724 tokens)
Code-switching detection rate	0.24% (1,915 phrases)
Orthographic contamination rate	3.66% (28,887 phrases)
Contaminant types	faux_ami: 31,805 / non_canonique: 10,092 / forbidden: 617
GlotLID confidence >= 0.95	99.7% (787,561 phrases)
DistilBERT HIGH_CONF (analyzed)	9,679 / 10,000
DistilBERT mismatch kab->tach	67 (0.67% of analyzed)

---

Column Structure

Column	Type	Description
id	string	Original Tatoeba identifier
text	string	Normalized Kabyle text (use for training)
raw_text	string	Original raw text (before normalization)
glot_lang	string	Language predicted by GlotLID v3
glot_conf	float	GlotLID confidence score
berber_label	string	DistilBERT label (kab / tach)
berber_score	float	DistilBERT confidence score
berber_status	string	HIGH_CONF / LOW_CONF / REJECT / NON_BERBER / SKIPPED
dominant_lang	string	Dominant language detected by MaskLID
dominant_ratio	float	Ratio of tokens in the dominant language
lang_distribution	string	JSON distribution of detected languages (token-level)

---

Quick Usage

from datasets import load_dataset

ds = load_dataset("boffire/tatoeba-kabyle-mono-cleaned", split="train")

# Example
print(ds[0]["text"])
# -> "Tecfiḍ fell-i ?"

# Iteration
for example in ds:
    print(example["text"])

---

Supplementary Files

The following files are available in the supplementary/ directory for full reproducibility and audit:

File	Description	Size
supplementary/tatoeba_kab_masklid.jsonl	Full 789,712 sentences with token-level CS annotations, contaminant lists, and all metadata	~539 MB
supplementary/tatoeba_kab_cs.conll	Code-switching sentences only, annotated in IOB2 format (CoNLL-U)	~214 KB
supplementary/tatoeba_kab_report.json	Machine-readable statistics and pipeline metadata	~1 KB

These files allow you to:

• Audit filtering decisions and apply custom thresholds
• Train or evaluate code-switching detection models on the IOB2-annotated subset
• Reproduce the statistics reported in the analysis report

---

Recommended Use Cases

• Continued pre-training of language models (LLM, mT5, etc.)
• Fine-tuning of Kabyle Sentence Transformers
• Tokenizer training for Kabyle-specialized SentencePiece / BPE
• Perplexity benchmark for evaluating the quality of new corpora
• Semantic search (FAISS indexing) and RAG

---

Known Limitations

1. Isolated sentence corpus: Tatoeba provides short utterances without discursive context. For LLM pre-training, this corpus should be complemented with longer documents (Wikipedia, forums, etc.).
2. Technical borrowings: Integrated technical terms in Kabyle (e.g., aselkim, asiggez) are retained as they belong to the contemporary lexicon. They do not constitute code-switching.
3. DistilBERT limited scope: Only the first 10,000 sentences were validated by DistilBERT; the remainder of the corpus is filtered by GlotLID + MaskLID only. Extrapolated mismatch rate: ~0.67% of the full corpus.
4. Standardized orthography: Normalization applies the INALCO 1996 / KIM inventory. Non-standard orthographic variants are corrected or discarded.
5. Short-sentence GlotLID noise: Sentences of 2-3 tokens containing proper names may trigger false language predictions. The allowlist fallback mitigates but does not eliminate this.
6. Static auxiliary vocabulary: Neologisms or proper names not present in the Hunspell dictionary or the auxiliary wordlist may be flagged as non-Kabyle tokens.

---

Related Resources

Resource	Link
Full analysis report	kabyle-specs.github.io/reports/report_tatoeba_kabyle.md
Raw Tatoeba dataset	tatoeba.org
Berber classifier	distilbert-kabyle-tachelhit-classifier-v2
Kabyle T5 tokenizer	kabyle-tokenizer-T5
Sentence Transformer	kabyle-sentence-transformer-mpnet
Hunspell dictionary	Imseɣti n tira n teqbaylit (Belkacem77, MIT)
Kabyle language specifications	kabyle-specs.github.io

---

Citation

If you use this dataset, please cite:

@dataset{tatoeba_kabyle_mono_cleaned_2026,
  author = {Mokraoui, Athmane},
  title = {Tatoeba Kabyle Mono Cleaned},
  year = {2026},
  publisher = {HuggingFace},
  url = {https://huggingface.co/datasets/boffire/tatoeba-kabyle-mono-cleaned}
}

---

Changelog

Version	Date	Change
1.0.0	2026-08-01	Initial release (756,774 pure sentences)
1.1.0	2026-08-01	Fixed data license (CC0 → CC BY 2.0 FR); added supplementary files (masklid JSONL, CS CoNLL-U, report JSON); expanded limitations and added supplementary files documentation

---

Dataset maintained by boffire. For questions or contributions, please open an issue on the repository.