magibu/wikipedia-40-langs

Wikipedia 40 Languages

A curated multilingual dataset of Wikipedia articles spanning 40 languages with 812,000 articles total. Designed for multilingual NLP research, language modeling, and cross-lingual transfer learning.

Dataset Summary

This dataset contains Wikipedia articles from 40 languages, sampled and split into train/validation/test sets with a consistent 10:3:1 ratio per language. English and Turkish are overrepresented (10x more samples) to support focused training scenarios, while the remaining 38 languages each contribute equally.

Property	Value
Total articles	812,000
Languages	40
Download size	2.78 GB
Dataset size	5.08 GB
License	CC BY-SA 4.0 (Wikipedia)

Splits

Split	Examples	Size
train	580,000	3.84 GB
validation	174,000	0.95 GB
test	58,000	0.30 GB

Split ratio is 10:3:1 (train:validation:test), applied consistently per language.

Features

Feature	Type	Description
lang	string	ISO 639 language code (e.g., en, tr, ja)
title	string	Wikipedia article title
text	string	Full article text content
url	string	Original Wikipedia article URL

Language Distribution

The dataset covers 40 languages. English (en) and Turkish (tr) each have 10x the samples of other languages.

Code	Language	Train	Validation	Test	Total
en	English	100,000	30,000	10,000	140,000
tr	Turkish	100,000	30,000	10,000	140,000
ar	Arabic	10,000	3,000	1,000	14,000
arz	Egyptian Arabic	10,000	3,000	1,000	14,000
bg	Bulgarian	10,000	3,000	1,000	14,000
ca	Catalan	10,000	3,000	1,000	14,000
ce	Chechen	10,000	3,000	1,000	14,000
ceb	Cebuano	10,000	3,000	1,000	14,000
cs	Czech	10,000	3,000	1,000	14,000
cy	Welsh	10,000	3,000	1,000	14,000
da	Danish	10,000	3,000	1,000	14,000
de	German	10,000	3,000	1,000	14,000
el	Greek	10,000	3,000	1,000	14,000
eo	Esperanto	10,000	3,000	1,000	14,000
es	Spanish	10,000	3,000	1,000	14,000
eu	Basque	10,000	3,000	1,000	14,000
fa	Persian	10,000	3,000	1,000	14,000
fi	Finnish	10,000	3,000	1,000	14,000
fr	French	10,000	3,000	1,000	14,000
he	Hebrew	10,000	3,000	1,000	14,000
hu	Hungarian	10,000	3,000	1,000	14,000
hy	Armenian	10,000	3,000	1,000	14,000
id	Indonesian	10,000	3,000	1,000	14,000
it	Italian	10,000	3,000	1,000	14,000
ja	Japanese	10,000	3,000	1,000	14,000
ko	Korean	10,000	3,000	1,000	14,000
ms	Malay	10,000	3,000	1,000	14,000
nl	Dutch	10,000	3,000	1,000	14,000
no	Norwegian	10,000	3,000	1,000	14,000
pl	Polish	10,000	3,000	1,000	14,000
pt	Portuguese	10,000	3,000	1,000	14,000
ro	Romanian	10,000	3,000	1,000	14,000
ru	Russian	10,000	3,000	1,000	14,000
sh	Serbo-Croatian	10,000	3,000	1,000	14,000
simple	Simple English	10,000	3,000	1,000	14,000
tt	Tatar	10,000	3,000	1,000	14,000
uz	Uzbek	10,000	3,000	1,000	14,000
vi	Vietnamese	10,000	3,000	1,000	14,000
war	Waray	10,000	3,000	1,000	14,000
zh	Chinese	10,000	3,000	1,000	14,000

Script Families Covered

The dataset spans multiple writing systems:

• Latin: en, tr, ca, ceb, cs, cy, da, de, eo, es, eu, fi, fr, hu, id, it, ms, nl, no, pl, pt, ro, sh, simple, uz, vi, war
• Cyrillic: bg, ce, ru, tt
• Arabic: ar, arz, fa
• CJK: ja, ko, zh
• Armenian: hy
• Greek: el
• Hebrew: he

Text Statistics (Train Split)

Statistic	Value
Min length	1 character
Max length	518,241 characters
Mean length	5,292 characters
Median length	1,585 characters
Std deviation	11,275 characters

The distribution is heavily right-skewed: 98.8% of articles are under 51,826 characters, with a long tail of very lengthy articles.

Usage

Loading with Hugging Face Datasets

from datasets import load_dataset

# Load the full dataset
dataset = load_dataset("alibayram/wikipedia-40-langs")

# Load a specific split
train = load_dataset("alibayram/wikipedia-40-langs", split="train")

# Filter by language
turkish_articles = train.filter(lambda x: x["lang"] == "tr")
english_articles = train.filter(lambda x: x["lang"] == "en")

# Stream for memory efficiency
streamed = load_dataset("alibayram/wikipedia-40-langs", split="train", streaming=True)

Example Data Point

{
  "lang": "en",
  "title": "Machine learning",
  "text": "Machine learning (ML) is a field of study in artificial intelligence...",
  "url": "https://en.wikipedia.org/wiki/Machine_learning"
}

Use Cases

• Multilingual language modeling: Pre-train or fine-tune language models across 40 languages
• Cross-lingual transfer learning: Evaluate how knowledge transfers between languages
• Machine translation: Use parallel topics across languages for indirect supervision
• Text classification: Train multilingual classifiers with language-balanced data
• Information retrieval: Build multilingual search and retrieval systems
• Script-diverse NLP: Study model behavior across Latin, Cyrillic, Arabic, CJK, and other scripts

Dataset Creation

Source

All articles are sourced from Wikipedia, the free encyclopedia. Each article's original URL is preserved in the url field for traceability.

Sampling Strategy

• English and Turkish: 100,000 articles each in the train split (oversampled for focused training)
• Other 38 languages: 10,000 articles each in the train split
• Split ratio: A consistent 10:3:1 ratio (train:validation:test) is maintained across all languages

Processing

Articles contain the full text content extracted from Wikipedia. The text field preserves the article body without markup.

Limitations and Biases

• Wikipedia coverage bias: Languages with larger Wikipedia editions may have higher-quality or more diverse articles. Smaller Wikipedias (e.g., Chechen, Waray) may contain more bot-generated or stub articles.
• Temporal snapshot: The dataset represents Wikipedia at a specific point in time and does not reflect subsequent edits.
• Content bias: Wikipedia has known biases in topic coverage (e.g., overrepresentation of Western-centric topics, gender imbalance in biographies).
• Uneven language oversampling: English and Turkish have 10x more samples, which may bias multilingual models toward these languages if not accounted for during training.
• No deduplication guarantees: Some articles may contain near-duplicate content (e.g., bot-generated geographic articles across languages).

License

The dataset inherits Wikipedia's Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0) license.

Citation

@dataset{wikipedia_40_langs,
  title={Wikipedia 40 Languages},
  author={Ali Bayram},
  year={2026},
  url={https://huggingface.co/datasets/alibayram/wikipedia-40-langs},
  license={CC BY-SA 4.0}
}