ArmWeb

ArmWeb is a curated Armenian news corpus for language-model pretraining: 4.37M documents / 11 GB / 3.3B Gemma-4 tokens (1.15B tokens in the 32K SentencePiece tokenizer used for the small-scale ablations) of deduplicated, decontaminated Eastern and Western Armenian text with document-level provenance (URL, outlet, topic, publication date). To our knowledge it is the first and largest open pretraining corpus built specifically for Armenian — existing Armenian text at scale exists only as slices of multilingual web crawls with no Armenian-specific curation.

Paper: From Zero to Hero: An Open LLM Ecosystem for Armenian (arXiv:2609.03350). Sister releases: COPA-AI/armstem and COPA-AI/arm-gemma-e4b.

Why another Armenian corpus?

Because the existing ones are contaminated and duplicated. Scanning the Armenian slices of public corpora against ten Armenian benchmarks (13-gram overlap, normalized text):

Corpus	Docs	Gemma-4 tokens	Benchmark-contaminated
ArmWeb (this release)	4.46M	3.3B	3.3% → removed
CulturaX-hy	2.96M	4.5B	7.9%
HPLT-v2-hy	3.60M	5.8B	10.9%
FineWeb-2-hy	1.76M	2.3B	17.4%

Token sizes re-tokenize each corpus as distributed with the Gemma-4 vocabulary (ArmWeb as released), so they are comparable: ArmWeb has the most documents but, being news, the least text of the three large corpora.

FineWeb-2's Armenian training split overlaps its own test split. Any evaluation of Armenian-trained models without decontamination is inflated. ArmWeb ships with its contamination removed and the scan code released.

ArmWeb is also complementary to the crawls, not redundant with them: in controlled 410M ablations it is ~10% better bits-per-byte on held-out news, and the union of ArmWeb + CulturaX beats every single corpus on the panel mean (0.532 vs 0.545 bpb for the best single corpus) and is within 0.007 bpb of the best on every panel — a lead that widens at 1.3B (see paper, Appendix).

Contents

• Source: a single-operator crawl of public Armenian news sites, 2011–2026, stored as a structured database (not raw HTML — near-zero boilerplate by construction).
• Splits: train (4.31M docs) / validation (20K) / test_iid (20K) / test_tail (20K; the final two months, for temporal generalization). Splits are stratified by outlet×month; deduplication ran globally before splitting, and three leakage gates are verified: zero exact cross-split collisions and near-duplicate rates ≤0.045% (val), 0.010% (test_iid), 0.005% (test_tail) run as hard assertions in the release build, and a third gate counts cross-split shared paragraphs of ≥13 tokens: 30 (val), 63 (test_iid), 10 (test_tail) per 20K-document split.
• Fields: id, text (NFC-normalized, otherwise verbatim), source (outlet), url, topic, post_date (publication), scrape_date (crawl), dedup_cluster_id, cluster_size (size of the near-duplicate cluster this document represents), split. Author names are not released.

Pipeline (fully documented in the paper)

Stage	Documents	Δ
Extracted	5,918,811	—
Language ID (GlotLID, hye/hyw)	5,817,343	−1.7%
Exact dedup (xxh128, keep-longest)	5,436,984	−6.5%
MinHash dedup (see below)	4,515,497	−16.9%
Repeated-paragraph removal	4,515,484	−0.0%
Splits held out (val/test_iid/test_tail, 20K each)	4,455,484	−60K
Benchmark decontamination (train side)	4,308,383	−3.3%

The dedup engine is cross-checked twice: the reference datasketch library, run with the identical recipe on a 1% sample, finds 449 in-sample duplicates of which the engine had already flagged 95.8%; NVIDIA NeMo-Curator at matched LSH geometry removes 23.5% vs our 22.4% on the full corpus with 99.2% per-document agreement on a matched-scope sample, disagreements confined to below-threshold borderline pairs (paper, Appendix). The identical recipe applied to a Russian sister collection from the same crawler removes only 0.9% of documents, so the high Armenian removal rate reflects real news syndication, not an aggressive pipeline.

MinHash: word 5-gram shingles on an aggressive NFKC "signature view" (Armenian ligature folding, punctuation stripping, digit-zeroing), 112 permutations, 14 bands × 8 rows (Jaccard ≈ 0.72). The lower-than-standard threshold is deliberate: news syndication produces true reprints at far lower Jaccard than web duplicates, and 0.8-threshold recipes miss most of them. Decontamination: 13-gram overlap against ten Armenian evaluation sets (Belebele, INCLUDE, HellaSwag-hy, MultiBLiMP, SIB-200, SynDARin, LR-Sum, the FineWeb-2-hy test split, hyWiki eval, FLORES-200), 147,101 documents removed. m-MMLU-hy and ARC-hy were not decontamination targets; a post-hoc scan found zero of the 4.31M training documents share any 13-gram with either benchmark.

License and attribution

The compilation (selection, cleaning, deduplication, metadata, splits) is released under ODC-BY 1.0. The crawl honored the sites' crawling rules, including robots.txt directives. The underlying article texts remain the copyright of their respective publishers; per-document url and source fields preserve attribution. Author bylines are removed; a PII scan (942 e-mail addresses in 782 documents, 8,971 documents with phone-number-like strings, overwhelmingly newsroom and institutional contact lines) ships with the corpus. The corpus is released for research use, consistent with text-and-data-mining practice for web-derived pretraining corpora (C4, OSCAR, FineWeb). If you are a rights holder and want content removed, open a discussion on this repository.

Citation

@article{arakelyan2026armweb,
  title  = {From Zero to Hero: An Open LLM Ecosystem for Armenian},
  author = {Arakelyan, Erik and Avetisyan, Khatun and Davtyan, Meri and Grigoryan, Heghine and Khachatryan, Nane and Shahsuvaryan, Hayk and Sergoyan, Henrik and Martirosyan, Vahan},
  year   = {2026},
  journal = {arXiv preprint arXiv:2609.03350}
}