Datasets:

kaan39
/

turkish-wikipedia-dataset-clean

configs:
  - config_name: default
    data_files:
      - split: train
        path: train.jsonl
license: cc-by-sa-4.0
language:
  - tr
tags:
  - turkish
  - wiki
  - wikipedia
  - turkish data
  - dataset
size_categories:
  - 100K<n<1M

Turkish Wikipedia Dataset

A cleaned and structured Turkish Wikipedia dataset designed for Turkish language model pretraining, continued pretraining, research, and NLP experiments.

The dataset consists of articles collected from the Turkish Wikipedia (tr.wikipedia.org) and processed into a machine-readable format while preserving important source metadata.

Dataset Summary

Language: Turkish (tr)
Source: Turkish Wikipedia
Domain: General knowledge / encyclopedia
Format: JSON / JSONL
Primary use: Turkish NLP and language-model training
Collection: Wikipedia API
Processing: Text extraction, normalization, cleaning, and metadata preservation

Each record represents a Wikipedia article.

Example

{
  "title": "1. Ordu (Osmanlı İmparatorluğu)",
  "text": "1. Ordu, Osmanlı Ordusu'nun ordularından biri.\n\nDüzen\nSavaş Düzeni, 1877\n...",
  "source": "https://tr.wikipedia.org/wiki/1._Ordu_%28Osmanl%C4%B1_%C4%B0mparatorlu%C4%9Fu%29",
  "page_id": 99567,
  "revision_id": 32655654,
  "timestamp": "2024-05-01T16:38:16Z"
}

Dataset Structure

Each JSON/JSONL record contains the following fields:

Field	Type	Description
`title`	string	Wikipedia article title
`text`	string	Cleaned article text
`source`	string	Original Turkish Wikipedia article URL
`page_id`	integer	Wikipedia page ID
`revision_id`	integer	Revision ID used for the dataset entry
`timestamp`	string	Timestamp associated with the source revision

Data Processing

The source articles are processed before being included in the dataset.

The processing pipeline is designed to:

Extract the main article content.
Remove unnecessary HTML and interface elements.
Normalize article text.
Preserve paragraph and section structure where possible.
Remove navigation and presentation-related content.
Preserve article titles.
Preserve the original Wikipedia source URL.
Preserve page and revision identifiers.
Preserve the source revision timestamp.
Produce UTF-8 encoded JSON/JSONL output.

The goal is to retain useful encyclopedic information while removing unnecessary webpage-related noise.

Dataset Format

The recommended training format is JSONL.

Each line represents one article:

{"title":"Article Title","text":"Article content...","source":"https://tr.wikipedia.org/wiki/...","page_id":123,"revision_id":456,"timestamp":"2025-01-01T00:00:00Z"}

This makes the dataset suitable for streaming and large-scale processing without requiring the complete dataset to be loaded into memory.

Intended Uses

This dataset may be useful for:

Turkish language model pretraining
Continued pretraining
Turkish NLP research
Text classification
Information retrieval
Semantic search
Question answering research
Turkish language understanding
Tokenizer evaluation
Text generation experiments
Dataset and preprocessing research

Out-of-Scope Uses

The dataset should not be treated as:

A source of guaranteed factual truth
A replacement for expert or primary sources
A curated medical, legal, financial, or scientific knowledge base
A dataset containing only verified or authoritative information

Wikipedia is collaboratively edited and may contain inaccuracies, incomplete information, disputed claims, outdated information, or biased coverage.

Users should independently verify important information against appropriate primary or authoritative sources.

Data Quality

The dataset is automatically processed and therefore may contain:

Incomplete articles
Historical or outdated information
Editorial inconsistencies
Formatting artifacts
Duplicate or near-duplicate information
Lists and tabular information represented as plain text
Articles with limited content

The dataset should therefore be considered a processed web/encyclopedic corpus, rather than a manually verified gold-standard dataset.

Source and Attribution

The underlying content originates from the Turkish Wikipedia:

https://tr.wikipedia.org/

Individual records retain their original article URL together with the corresponding Wikipedia page ID and revision ID whenever available.

Users redistributing or creating derivative datasets from this corpus should comply with the licensing requirements of the original Wikipedia content.

License

The textual content of Wikipedia is available under the Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0) license, subject to the applicable terms and conditions.

https://creativecommons.org/licenses/by-sa/4.0/

When redistributing this dataset or a derivative dataset, users should review the current Wikipedia and Wikimedia licensing requirements and provide appropriate attribution.

The dataset creator does not claim ownership of the underlying Wikipedia article text.

Citation

If you use this processed dataset in research or another public project, please cite this dataset and the original source.

Dataset

@misc{turkish_wikipedia_dataset,
  title        = {Turkish Wikipedia Dataset},
  author       = {Dataset Contributors},
  year         = {2026},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/YOUR_USERNAME/YOUR_DATASET}
}

Original Source

@misc{wikipedia,
  title        = {Wikipedia},
  publisher    = {Wikimedia Foundation},
  url          = {https://www.wikipedia.org/}
}

Disclaimer

This dataset is provided for research and educational purposes.

The dataset creator is not responsible for errors, omissions, outdated information, or claims contained in the underlying Wikipedia articles.

The presence of an article or statement in this dataset does not constitute endorsement of its content.

Versioning

Dataset releases should be versioned independently from the source Wikipedia revisions.

Example:

train0001
train0002
train0003

Each release may contain different articles or revisions depending on the collection date and preprocessing pipeline.

The revision_id and timestamp fields should be used to identify the exact source revision represented by an individual record.

Acknowledgements

This dataset would not be possible without the contributors who create, edit, maintain, and improve Turkish Wikipedia.

Special thanks to the Turkish Wikipedia community and Wikimedia contributors.

Source: Turkish Wikipedia (tr.wikipedia.org) Language: Turkish Format: JSON / JSONL License of source content: CC BY-SA 4.0