Hugging Face's logo

Datasets:
eckendoerffer
/
wikipedia_fr

Tasks:
Text Generation
Modalities:
Text
Formats:
text
Languages:
French
Size:
1M - 10M
Tags:
wikipedia
wiki
fr.wikipedia.org
Libraries:
Datasets
License:
Dataset card Viewer Files Community
wikipedia_fr / README.md
eckendoerffer's picture
eckendoerffer
Update README.md
a7ff7a5
preview code
|
raw
history blame contribute delete
3.03 kB
metadata
license: cc-by-sa-3.0
task_categories:
  - text-generation
language:
  - fr
tags:
  - wikipedia
  - wiki
  - fr.wikipedia.org
size_categories:
  - 1M<n<10M

French Wikipedia Dataset

Overview

This dataset is a curated collection of approximately 1.1 million French Wikipedia articles, scraped directly from the official French Wikipedia site on September 24, 2023. There are already numerous datasets for Wikipedia, including the official one with Wikipedia's dump. Unfortunately, the text for the French version of this dataset is incomplete, lacking many elements like dates and locations. As the saying goes, "garbage in, garbage out."

Format

  • Type: Text
  • File Extension: .txt

Structure

The dataset is divided into the following splits:

  • train.txt: 3.45 GB - 1,810,000 rows - 90%
  • test.txt : 192 MB - 100,575 rows - 5%
  • valid.txt: 192 MB - 100,575 rows - 5%

Each article in the dataset exceeds 1400 characters in length.

Data Cleaning and Preprocessing

The following elements have been excluded from the dataset:

  • H1 - H4 Headings
  • Lists
  • Tables
  • Sources and References
  • Info box
  • Banners
  • LaTeX code

The text has been standardized for consistent formatting and line length. Additionally, the dataset has been filtered using the langid library to include only text in French. Some quotations or short terms in other languages, including non-Latin languages, may still be present.

Exploring the Dataset

You can use the explore_dataset.py script to explore the dataset by randomly displaying a certain number of lines from it. The script creates and saves an index based on the line breaks, enabling faster data retrieval and display.

Additional Information

This dataset is a subset of a larger 10GB French dataset, which also contains several thousand books and theses in French, as well as several hundred thousand Francophone news articles.

WIKIPEDIA EXTRACT

Inside the /extract_wiki/ directory, you'll find Python scripts used to extract text to compile this dataset.

Requirements: 

pip install datasets aiohttp aiofiles beautifulsoup4 langid

Scripts:

  1. 1_extract_link.py

    python 1_extract_link.py

    Script to download the Wikipedia dataset from Hugging Face, extract URLs, and save them to a text file for further processing.

  2. 2_extract_content.py

    python 2_extract_content.py

    This script retrieves the source code of Wikipedia pages based on URLs found in a text file. Instead of saving the entire HTML of the page, it trims the content, focusing on the main article section, thereby limiting the size of each record.

  3. 3_extract_txt.py

    python 3_extract_txt.py

    This script extracts the text from the HTML pages and conducts tests to filter the content that should be retained or excluded. This includes language checks, special characters, numbers, etc.