DATASHEET.md

Datasheet — Azerbaijani Pretraining Corpus (v0)

A cleaned, deduplicated, PII-redacted ~1.0 billion token Latin-script Azerbaijani corpus for language-model pretraining, assembled from open multilingual web + encyclopedic sources with a reproducible datatrove pipeline. Structured as a Datasheet for Datasets (Gebru et al.).

Built as Step 2 of the Azerbaijani LLM stack (see CLAUDE.md); the pipeline transfers to other Turkic languages. Token counts use the project's az_unigram_32k SentencePiece tokenizer.

---

At a glance

Language	Azerbaijani (az / azj), Latin script only
Documents	1,711,442
Tokens	1,003,687,462 (az_unigram_32k; uint16) — train 998.6M / val 5.1M
Uncompressed text	~5.37 GB
Split	99.5% train / 0.5% val (document-level, seed 42)
Format	nanoGPT uint16 bins (train.bin, val.bin) + meta.json; `<
License	mixed per source (see Licensing) — redistribution as a derived corpus
Pipeline	data/ in this repo (fully reproducible)

---

Composition

Sources & filtering funnel

source	HF repo (config)	docs after LID+script	after quality/PII	kept	final (post-dedup)	share
CulturaX	uonlp/CulturaX (az)	983,451	883,386	89.8%	864,539	50.5%
HPLT v2	HPLT/HPLT2.0_cleaned (azj_Latn)	988,573	904,204	91.5%	697,071	40.7%
mC4	allenai/c4 (az)	414,310	308,400	74.4%	145,512	8.5%
Wikipedia	wikimedia/wikipedia (20231101.az)	7,512	5,656	75.3%	4,320	0.25%
Total		2,393,846	2,101,646		1,711,442	100%

Cross-source MinHash dedup removed 390,204 docs (18.6%). The largest cut was mC4: 308,400 → 145,512 (−53%) — CulturaX is derived from mC4+OSCAR, so most of our mC4 slice was duplicate CulturaX content and was correctly removed. HPLT (an independent crawl) overlapped little and mostly survived.

Provenance note: CulturaX rows carry their own source field ("mC4"/"OSCAR"); in the packed metadata CulturaX documents therefore appear under their CulturaX-internal provenance rather than the label culturax. The counts above use the intended source identity.

Register

Predominantly web text (CulturaX, HPLT, mC4 ≈ 99.7%) with a small encyclopedic slice (Wikipedia). This is a web-register corpus; it is not balanced for genre, domain, or dialect.

What each instance is

One instance = one document (a web page or article), as plain text. No labels, no annotations — this is an unlabeled pretraining corpus.

---

Languages & script decision

Latin-script Modern (North) Azerbaijani only. Cyrillic (U+0400–04FF) and Perso-Arabic (U+0600–06FF) text is dropped at the script-filter stage; we do not transliterate. Rationale (see repo README): Republic-of-Azerbaijan Azerbaijani has been officially Latin since ~2001 and all current production text is Latin. Cyrillic (legacy) and Perso-Arabic (South Azerbaijani) are different orthographies/distributions and out of scope for v0. Casing is preserved, including the semantically meaningful dotted/dotless-i distinction (i/İ vs ı/I).

---

Collection process

Documents were streamed from the public HuggingFace dataset repos above (no re-crawling). CulturaX is acceptance-gated (terms accepted + token); the others are open. We did not collect from individuals; all sources are existing public web/encyclopedic datasets. Collection date: 2026-06. OSCAR-2301 and MADLAD-400 were attempted but excluded from v0 (OSCAR gating pending; MADLAD config resolution impractical here).

---

Preprocessing / cleaning

Each document passed, in order (datatrove blocks unless noted; the Azerbaijani-specific pieces are custom):

1. Script filter (custom) — keep documents ≥85% Latin letters, <5% Cyrillic/Arabic, ≥200 chars.
2. Language ID — fastText lid.176, keep az at ≥0.65 confidence (distinguishes Azerbaijani from Turkish cleanly: az→az @1.000, not tr).
3. Gopher repetition filter — drop documents dominated by repeated lines / n-grams.
4. Gopher quality filter — length, mean word length, symbol ratios, and an Azerbaijani stop-word presence check (custom ~50-word list; datatrove's default is English).
5. C4 quality filter — per-line web cruft (too-short lines, JS, lorem-ipsum, stray braces).
6. PII redaction — emails and public IP addresses replaced with placeholders.
7. MinHash near-duplicate removal — across all sources jointly (5-grams, 14×8 LSH, 64-bit), tokenized with a custom Azerbaijani whitespace/regex word tokenizer (no spaCy).
8. Tokenize + pack — az_unigram_32k, document-shuffled (buffer 100k, seed 42), uint16 bins.

Word/sentence tokenization for the filters uses a custom regex tokenizer faithful to Gopher's original whitespace word-counting and correct for agglutinative Latin Azerbaijani.

---

Uses

• Intended: pretraining / continued-pretraining of Azerbaijani (or multilingual) language models; tokenizer evaluation; corpus linguistics; a base to mix with other languages.
• Out of scope: anything requiring labels (it has none); high-stakes or factual-authority use (it is unfiltered-for-truth web text); Cyrillic/Perso-Arabic Azerbaijani (excluded); dialectal balance.

---

Limitations & biases (honest)

• Web-register skew. ~99.7% web text → boilerplate, SEO, commercial, and news-heavy distributions; Wikipedia is the only encyclopedic counterweight (0.25%). Not genre-balanced.
• Size. ~1.0B tokens is solid for a low-resource language but ~8 tokens/param for a 124M model (Chinchilla-optimal ≈ 20) — expect to overtrain or use as continued-pretraining data.
• Quality is heuristic, not human-reviewed. Gopher/C4 rules catch gross junk; subtle low-quality, toxicity, social bias, and factual errors in web text are not removed. No toxicity classifier was run.
• Lineage overlap. CulturaX/mC4 share mC4 lineage; dedup removes near-duplicates but exact provenance weighting is approximate. CulturaX dominates (50.5%), so its biases dominate.
• PII redaction is regex-level (emails/IPs) — not a guarantee; names and other PII in public web text may remain.
• Latin-only by design — excludes Cyrillic and South Azerbaijani (Perso-Arabic) speakers' text.

---

Distribution & licensing

This is a derived corpus; per-source upstream terms apply:

source	license
CulturaX	ODC-BY + per-document terms
HPLT v2	per HPLT release terms (CC0 metadata)
mC4 / C4	ODC-BY
Wikipedia	CC-BY-SA 4.0

The CC-BY-SA (Wikipedia) component implies share-alike obligations on redistribution. Consult each upstream dataset's terms before redistributing. We redistribute only as documented derived artifacts/pointers.

---

Reproduction

# (gated CulturaX needs: hf auth login + accept terms)
.venv-data/bin/python data/download/fetch.py --sources culturax,hplt,mc4,wikipedia --max-retries 4
.venv-data/bin/python data/filter/clean.py  --sources all
.venv-data/bin/python data/filter/dedup.py
.venv-data/bin/python data/pack/tokenize_pack.py
.venv-data/bin/python data/stats/report.py

Exact configs, seeds, and the dedup tasks=1 determinism fix are in data/ and experiments/RUN_LOG.md. Pipeline payloads are gitignored; counts/decisions are committed.

Maintenance

v0, 2026-06. Planned: add OSCAR-2301 (gating pending) and MADLAD-400; broaden register; build a held-out eval that spans registers. Contact: project maintainer (see repo).

Citation

A formal citation will accompany the HuggingFace release. For now cite the repository and the upstream datasets (CulturaX, HPLT v2, mC4/C4, Wikipedia).