open-index/open-wikipedia-markdown

Open Wikipedia (Markdown)

Every Wikipedia article converted to clean Markdown, 139.4K articles across 1 languages

What is it?

This dataset contains every article from every language edition of Wikipedia, converted from raw MediaWiki markup into clean, readable Markdown. Headings, bold, italic, code blocks, and internal links are all preserved as proper Markdown syntax, while templates, infoboxes, references, tables, categories, and other noise are stripped away.

The source data comes from the official Wikimedia database dumps. Each language's full XML export is streamed, parsed, and converted article by article. The results are stored as sharded Apache Parquet files with Zstandard compression, organized by language. Each language gets its own directory under data/, and each shard holds up to 500,000 articles.

This is the Markdown variant of the Open Wikipedia collection. If you need plain text with all formatting removed, see open-index/open-wikipedia-text. If you need the original MediaWiki source markup, see open-index/open-wikipedia.

139.4K articles | 1 languages | Last updated: 2026-04-03 | License: CC BY-SA 4.0

What is being released?

The dataset is organized as one directory per language, with sharded Parquet files inside each:

data/
  en/en-00000.parquet      English, shard 0
     en-00001.parquet      English, shard 1
     ...
  de/de-00000.parquet      German
  fr/fr-00000.parquet      French
  es/es-00000.parquet      Spanish
  ja/ja-00000.parquet      Japanese
  ...
  la/la-00000.parquet      Latin

Each Parquet file contains up to 500,000 rows. Languages with fewer articles than that fit in a single shard. Larger languages like English, German, and French are split across multiple shards. All files use Zstandard compression.

How to download and use this dataset

Using DuckDB

DuckDB can read Parquet files directly from Hugging Face without downloading anything first. This is the fastest way to explore the data.

-- Count articles per language
SELECT lang, COUNT(*) as articles
FROM read_parquet('hf://datasets/open-index/open-wikipedia-markdown/data/*/*.parquet')
GROUP BY lang
ORDER BY articles DESC;

-- Search for articles about a topic across all languages
SELECT title, lang, length, url
FROM read_parquet('hf://datasets/open-index/open-wikipedia-markdown/data/*/*.parquet')
WHERE markdown ILIKE '%machine learning%'
ORDER BY length DESC
LIMIT 20;

-- Article length distribution for English Wikipedia
SELECT
    percentile_disc(0.25) WITHIN GROUP (ORDER BY length) AS p25,
    percentile_disc(0.50) WITHIN GROUP (ORDER BY length) AS p50,
    percentile_disc(0.75) WITHIN GROUP (ORDER BY length) AS p75,
    percentile_disc(0.90) WITHIN GROUP (ORDER BY length) AS p90,
    percentile_disc(0.99) WITHIN GROUP (ORDER BY length) AS p99,
    AVG(length)::INT AS avg_length,
    MAX(length) AS max_length
FROM read_parquet('hf://datasets/open-index/open-wikipedia-markdown/data/en/*.parquet');

-- Find the longest articles in each language
SELECT lang, title, length, url
FROM (
    SELECT *, ROW_NUMBER() OVER (PARTITION BY lang ORDER BY length DESC) AS rn
    FROM read_parquet('hf://datasets/open-index/open-wikipedia-markdown/data/*/*.parquet')
)
WHERE rn = 1
ORDER BY length DESC
LIMIT 20;

-- How many articles contain code blocks?
SELECT lang, COUNT(*) AS articles_with_code, COUNT(*) * 100.0 / SUM(COUNT(*)) OVER () AS pct
FROM read_parquet('hf://datasets/open-index/open-wikipedia-markdown/data/*/*.parquet')
WHERE markdown LIKE '%```%'
GROUP BY lang
ORDER BY articles_with_code DESC
LIMIT 15;

Using datasets

from datasets import load_dataset

# Load English Wikipedia
ds = load_dataset("open-index/open-wikipedia-markdown", "en")
print(ds["train"][0]["title"])
print(ds["train"][0]["markdown"][:500])

# Stream the full dataset without downloading everything first
ds = load_dataset("open-index/open-wikipedia-markdown", "en", split="train", streaming=True)
for item in ds:
    print(item["title"], item["length"])

# Load a specific language
ds = load_dataset("open-index/open-wikipedia-markdown", "de")
print(f"German articles: {len(ds['train']):,}")

Using huggingface_hub

from huggingface_hub import snapshot_download

# Download only English
snapshot_download(
    "open-index/open-wikipedia-markdown",
    repo_type="dataset",
    local_dir="./wiki-md/",
    allow_patterns="data/en/*",
)

For faster downloads, install pip install huggingface_hub[hf_transfer] and set HF_HUB_ENABLE_HF_TRANSFER=1.

Using the CLI

# Download a single language
huggingface-cli download open-index/open-wikipedia-markdown \
    --include "data/la/*" \
    --repo-type dataset --local-dir ./wiki-md/

Using Polars

import polars as pl

df = pl.read_parquet("data/en/*.parquet")
print(f"English articles: {len(df):,}")
print(f"Total Markdown text: {df['length'].sum() / 1e9:.1f} GB")
print(df.select("title", "length").describe())

Dataset statistics

Languages

Language	Code	Articles	Shards
Latin	la	139.4K	1

Schema

Every Parquet file shares the same schema:

Column	Type	Description
id	int64	Wikipedia page ID, unique within each language edition
title	string	Article title as it appears on Wikipedia
markdown	string	Full article body converted from wikitext to Markdown
url	string	Direct URL to the Wikipedia article
lang	string	ISO 639 language code (e.g. en, de, fr, ja)
length	int32	Markdown body length in bytes
timestamp	string	Last revision timestamp in ISO 8601 format

Example instance

Here is an example row from the English partition, showing a converted article:

{
  "id": 12,
  "title": "Anarchism",
  "markdown": "# Anarchism\n\n**Anarchism** is a political philosophy and movement that is against all forms of authority...",
  "url": "https://en.wikipedia.org/wiki/Anarchism",
  "lang": "en",
  "length": 87453,
  "timestamp": "2025-12-15T08:22:01Z"
}

The markdown field contains the full article text with Markdown formatting. Internal wiki links are converted to full Wikipedia URLs, so [[United States|US]] becomes [US](https://en.wikipedia.org/wiki/United_States).

Wikitext to Markdown conversion

The conversion handles the most common MediaWiki syntax elements and maps them to their Markdown equivalents:

MediaWiki syntax	Markdown output
== Heading ==	## Heading
=== Subheading ===	### Subheading
'''bold'''	**bold**
''italic''	*italic*
[[Page|Text]]	[Text](https://lang.wikipedia.org/wiki/Page)
[https://example.com text]	[text](https://example.com)
<syntaxhighlight lang="python">	```python ```
<code>x</code>	`x`
<pre>block</pre>	```\nblock\n```

What gets stripped

The following elements are removed during conversion to produce clean, readable text:

Element	Handling
{{templates}}	Removed entirely, including Infobox, Navbox, Taxobox, and all other templates
{{Infobox ...}}	Removed, including nested template parameters
`{	tables
<ref> citations	Removed, including named references
[[File:]] / [[Image:]]	Removed, including thumbnails and captions
[[Category:]]	Removed
<!-- comments -->	Removed
Interwiki links	Removed
Magic words	__NOTOC__, __FORCETOC__, and similar directives are removed

The goal is to preserve the article's readable content and structure while removing everything that only makes sense in the context of the MediaWiki rendering engine.

How it works

The pipeline processes all 1 Wikipedia language editions through the following steps:

1. Download. The latest {lang}wiki-latest-pages-articles.xml.bz2 dump is streamed from dumps.wikimedia.org. Downloads support HTTP range resumption, so interrupted transfers pick up where they left off.

2. Parse. A streaming XML parser processes the bzip2-compressed dump without extracting it to disk. Only namespace-0 pages (articles) are kept. Redirects, talk pages, user pages, and all other namespaces are skipped.

3. Convert. Each article's wikitext is converted to Markdown through a series of regex-based transformations. Templates are stripped with up to 5 nesting passes to handle deeply nested constructs. Internal wiki links are resolved to full Wikipedia URLs for the appropriate language edition.

4. Filter. Articles shorter than 100 bytes after conversion are excluded. This removes stubs, disambiguation pages, and other pages with minimal content.

5. Shard. Articles are written to Zstandard-compressed Parquet files, approximately 500,000 rows per shard. Multiple languages are processed in parallel using a worker pool.

6. Publish. Each language's shards are committed to this Hugging Face repository as they complete. Commit messages include article counts, shard counts, and file sizes for auditability.

Considerations

Why Markdown instead of plain text?

Plain text is sufficient for many NLP tasks, but it loses document structure. Markdown preserves headings, bold, italic, code blocks, and links, which makes it better suited for:

• Language model training where the model should understand document structure
• Retrieval-augmented generation (RAG) where chunking by heading sections produces more coherent results
• Knowledge graph construction where preserved links encode relationships between concepts

If you do not need formatting, the plain text variant is smaller and simpler.

Known limitations

• Conversion is regex-based, not a full parser. Some complex wikitext constructs (deeply nested tables inside templates, parser functions, Lua module output) may not convert perfectly. The vast majority of articles convert cleanly, but edge cases exist.
• Templates are stripped, not expanded. Infoboxes, navigation boxes, and other templates are removed entirely rather than expanded to their rendered output. This means some structured data that appears in rendered Wikipedia pages is not present in this dataset.
• One snapshot in time. This dataset represents a single snapshot of each language's dump. It does not track edit history or article revisions.
• Dump availability varies. Not all language editions have their dumps available at all times. Languages whose dumps fail to download are skipped and will be included in future updates.

Related datasets

• open-index/open-wikipedia-text - Same articles as pure plain text with all formatting removed. Smaller files, better for embeddings and classification.
• open-index/open-wikipedia - Same articles in original MediaWiki wikitext markup. Use this if you need templates, tables, references, and other source elements.

Thanks

The content in this dataset was written by millions of Wikipedia editors worldwide and is hosted by the Wikimedia Foundation. The raw data comes from the Wikimedia database dumps, which the Foundation makes freely available for download.

Wikipedia is one of humanity's greatest collaborative achievements. All credit for the content goes to the volunteer editors who write, review, and maintain it.

This dataset is an independent conversion and is not affiliated with or endorsed by the Wikimedia Foundation.

Licensing

Wikipedia content is released under the Creative Commons Attribution-ShareAlike 4.0 International License (CC BY-SA 4.0). This dataset inherits that license. If you redistribute or build upon this data, you must give appropriate credit and share your contributions under the same license.

Citation

@dataset{open_wikipedia_markdown,
  title     = {Open Wikipedia (Markdown)},
  author    = {Open Index},
  year      = {2026},
  url       = {https://huggingface.co/datasets/open-index/open-wikipedia-markdown},
  license   = {CC BY-SA 4.0},
  publisher = {Hugging Face}
}

Last updated: 2026-04-03