README.md

metadata

license: cc-by-sa-4.0
language:
  - so
size_categories:
  - 100K<n<1M
task_categories:
  - text-generation
  - fill-mask
pretty_name: SomaliWeb v1
configs:
  - config_name: default
    data_files:
      - split: train
        path: train-*.parquet
      - split: validation
        path: validation.parquet
tags:
  - somali
  - low-resource
  - pretraining
  - web-corpus
  - deduplicated
  - quality-filtered
  - arxiv:2605.18232

SomaliWeb v1 — Quality-filtered Somali web corpus

📄 Paper: arXiv:2605.18232 — SomaliWeb v1: A Quality-Filtered Somali Web Corpus with a Matched Tokenizer and a Public Language-Identification Benchmark

💻 Construction pipeline (MIT): github.com/khaledyusuf44/somali-corpus

SomaliWeb v1 is a cleaned, deduplicated, and quality-filtered Somali-language web corpus of ~303 million tokens (819,322 documents), built by aggregating three public Somali-heavy web distributions (HPLT v2, CC100, Somali Wikipedia) and passing them through a reproducible six-stage pipeline.

This is, to our knowledge, the first dedicated and versioned Somali-only pretraining corpus released on Hugging Face with a complete dataset card, a matched tokenizer, and a public language-identification benchmark. The full audit-trail, equations, and findings are documented in the companion paper (arXiv:2605.18232).

Dataset summary

• Language: Somali (Standard Somali, Latin script). No Maay Maay detected in the final distribution.
• Token count: ~303M approximate tokens (words × 1.3), 233M whitespace-separated words.
• Document count: 819,322 (train: 778,355 · validation: 40,967, 95/5 split).
• Script: Latin.
• Period of source content: ~2019–2024, inherited from HPLT v2 and CC100.
• License: CC-BY-SA 4.0 (see Licensing below; sources have compatible terms).

Why this dataset

Somali is a low-resource language with ~25 million speakers worldwide. Existing Somali text appears inside multilingual dumps (HPLT, CC100, mC4, OSCAR) but no standalone, named, versioned Somali pretraining corpus has been published with a documented construction pipeline. SomaliWeb v1 fills that gap and ships with:

• A reproducible pipeline anyone can re-run and audit.
• A comparison baseline tokenizer trained on HPLT-raw, released alongside.
• An evaluation against FLORES-200 Somali showing measurable fertility improvements over general-purpose tokenizers.

Dataset structure

Data fields

Each row in the train shards (train-*.parquet) and validation.parquet has the schema:

• id (string): source-prefixed unique identifier (e.g. hplt2-so_12345).
• text (string): the document body, cleaned and length-filtered.
• source (string): one of hplt2-so, cc100-so, wikipedia-so.
• n_words (int): whitespace-separated word count after cleaning.
• quality_score (float): Phase-5 quality score (0–1, higher = cleaner Somali).
• lang_conf (float): langdetect top-1 confidence for Somali.
• dialect_tag (string, optional): GlotLID v3 dialect tag (almost always som_Latn).

Data splits

Split	Documents	Approx tokens
train (95%)	778,355	~288M
validation (5%)	40,967	~15M

Source composition (of the combined train + validation)

Source	Documents	Fraction
HPLT v2 som_Latn	582,257	71.07%
CC100-so (statmt.org)	233,394	28.49%
Somali Wikipedia (dump 2023-11-01)	3,671	0.45%

Construction pipeline

Full details + intermediate metrics in the companion GitHub repository's reports/ directory. The pipeline is six phases:

1. Merge + byte-exact dedup. Combine the three sources under a unified schema; drop documents whose SHA-256 of lowercase(whitespace-collapsed(text)) is already seen (first-seen wins). Removed 189,692 duplicates (13.83%), of which 166,628 (17.3%) came from within HPLT v2's own "cleaned" distribution, demonstrating that HPLT's dedup is not byte-exact.

2. Clean + normalize. Mojibake fix via ftfy, whitespace collapse, repeated-char-run collapse, minimum-length filter (≥ 50 words). ftfy fixed mojibake in 615,314 documents (52.04% of input) — notably 56.1% of HPLT v2's documents contained fixable encoding artifacts.

3. LID verification. langdetect (seeded) on every surviving document; keep only top-1 = so with confidence ≥ 0.50. Drop rate 0.21% — HPLT/CC100 already LID-filtered solidly. GlotLID v3 second pass tagged dialect: som_Latn 99.96%, ymm_Latn (Maay Maay) 0.00%.

4. MinHash near-duplicate detection. Word-3-gram shingles hashed inline to 31-bit ints, 64 MinHash signatures, (16, 4) LSH banding (s* ≈ 0.50), τ = 0.80 Jaccard verification, union-find clustering, keep-longest rule. Removed 113,896 near-duplicates (10.57%). Largest cluster: 75 copies of one article — HPLT missed it.

5. Quality filter. Character 5-gram coverage against a clean Somali Wikipedia seed (2,221 articles ≥ 200 words, 828K unique 5-grams). Dropped bottom 15% by coverage score (threshold 0.9029). Filter catches template-like listings and encoding-degraded content; a known false-positive rate applies to news with heavy proper-noun density.

6. Release structuring. Shuffled (seed 0), 95/5 train/val split, metadata consolidated.

Evaluation

Tokenizer fertility on FLORES-200 Somali devtest (1,012 held-out sentences)

Tokenizer	Training corpus	Vocab	Tokens	Fertility ↓
SomaliWeb-v1 (this dataset)	350M tokens (cleaned)	16K	35,867	1.538
HPLT-raw	505M tokens (raw HPLT v2)	16K	35,854	1.537
GPT-4 cl100k_base	proprietary	100K	60,010	2.573

Two results:

• SomaliWeb v1 matches HPLT-raw's tokenizer fertility at 30% smaller training corpus — our dedup + quality filter preserves tokenizer quality.
• SomaliWeb v1's tokenizer is 40.2% more token-efficient than GPT-4's cl100k_base on Somali text. The "tokenization tax" that penalizes low-resource languages is concretely measurable here.

The companion tokenizer is published alongside as tokenizer_somaliweb.json (HF tokenizers JSON format).

Limitations and considerations

• Source inheritance. The upstream sources (HPLT v2, CC100, Somali Wikipedia) each have their own coverage biases. HPLT and CC100 are CC-derived and therefore dominated by whatever Somali content happened to be on the public web during ~2019–2024.
• No Maay Maay. GlotLID v3 tagged 0 documents as ymm_Latn in the final distribution. Either the upstream distributions filtered it out, or GlotLID undercounts Maay on web text — either way, this corpus should be described as Standard Somali only. Future versions should source Maay Maay content separately.
• Quality filter false positives. The char-5-gram filter can drop real news articles heavy in proper nouns because the seed is predominantly Wikipedia encyclopedic prose. v2 should use a mixture seed or a trained classifier.
• Known encoding artifacts caught, not created. 52% of docs had mojibake fixed by ftfy. We did not find evidence of residual mojibake, but this hasn't been exhaustively audited.
• Dialect scope. Standard Somali only.
• PII. We did not run a dedicated PII-removal pass. Upstream sources may retain names, phone numbers, or emails embedded in news content. Empirical scan of the released splits: ~7.9% of documents contain at least one email-shaped string (mostly newsroom contact addresses such as editor@…, support@…, but also occasional personal gmail.com / hotmail.com addresses). Presidio does not cover Somali; a Somali-specific PII filter is planned for v2. Users building consumer-facing applications must apply additional PII handling.
• Copyright. Sources are public web documents. Somali news sites are the largest contributors via HPLT/CC100. Users should consider downstream licensing obligations for their specific use case.
• Bias. The corpus inherits any biases of the upstream crawls — the Somali internet skews toward diaspora, news, politics, and religion. Under-represented registers include conversational speech, technical writing, and long-form fiction.

Usage

Load with datasets

from datasets import load_dataset

ds = load_dataset("khaledyusuf44/somaliweb-v1")
print(ds)
# DatasetDict({'train': Dataset(num_rows=778355), 'validation': Dataset(num_rows=40967)})

sample = ds["train"][0]
# {'id': 'hplt2-so_12345', 'text': '...', 'source': 'hplt2-so', ...}

Load the companion tokenizer

from tokenizers import Tokenizer
tok = Tokenizer.from_file("tokenizer_somaliweb.json")
print(tok.encode("Soomaaliya waa dal ku yaal geeska Afrika.").ids)

Intended uses

• Somali LLM pretraining (the primary use case).
• Somali tokenizer training and evaluation.
• Low-resource NLP research.
• Somali downstream task fine-tuning (classification, NER, translation) — as a base for domain adaptation.

Out-of-scope uses

• Consumer-facing applications without additional PII filtering.
• Claims of representing all Somali dialects. This corpus is Standard Somali only.
• Evaluation tasks that require contamination control. Source documents predate construction and may overlap with FLORES-200 or other public Somali evaluation sets. We recommend a contamination audit before using this corpus to train models you intend to evaluate on those benchmarks.

Dataset Curators

Khalid Yusuf Dahir (khaledyusuf44 on GitHub and Hugging Face).

Licensing

Released under CC-BY-SA 4.0, inheriting the most restrictive license among the upstream sources:

• HPLT v2: CC0 (permissive).
• CC100: public domain / MIT-like (permissive).
• Somali Wikipedia: CC-BY-SA 3.0 (requires attribution + share-alike).

Users of SomaliWeb v1 must honor CC-BY-SA 4.0 (attribution + share-alike). Citations appreciated.

Citation

If you use SomaliWeb v1, the matched tokenizer, the LID benchmark, or the construction pipeline, please cite the companion paper:

@article{dahir2026somaliweb,
  title         = {SomaliWeb v1: A Quality-Filtered Somali Web Corpus with a Matched Tokenizer and a Public Language-Identification Benchmark},
  author        = {Dahir, Khalid Yusuf},
  year          = {2026},
  eprint        = {2605.18232},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2605.18232}
}

Acknowledgments

This dataset builds on:

• HPLT v2 (University of Helsinki / Turku / Edinburgh consortium) — the primary CC-derived source.
• CC100 (Wenzek et al., 2020) — the secondary CC-derived source distributed via statmt.org.
• Somali Wikipedia contributors — the clean anchor and quality-filter seed.
• Common Crawl — the underlying web archive all three upstream corpora were built from.
• FLORES-200 (NLLB Team et al., 2022) — held-out evaluation set.
• GlotLID (Kargaran et al., 2023) — dialect tagging.

Changelog

• v1.0 (2026-04-23): Initial release.
• v1.0.1 (2026-05-19): Added arXiv:2605.18232 paper citation and reference.