Hugging Face's logo

Datasets:
recursal
/
SuperWikiNEXT-32B

Tasks:
Text Generation
Fill-Mask
Modalities:
Text
Formats:
json
Sub-tasks:
language-modeling
masked-language-modeling
Languages:
Afrikaans
Arabic
Asturian
Size:
100K - 1M
Libraries:
Datasets
Dask
License:
Dataset card Viewer Files Community
SuperWikiNEXT-32B / DOCS.md
KaraKaraWitch's picture
KaraKaraWitch
Initial Commit
d93fd32
preview code
|
raw
history blame
5.15 kB

Documentation

This contains some really quick docs and notes on filtering with SuperWIKI NEXT.

wikipedia_soup.py

...Is the main class that handles the bulk of the filtering.

Each filter has code documentation to explain what each function generally does. So I'd suggest you to read those instead.

Usage for wikipedia_soup.py

probably the most important bit.

wikipedia_soup takes in *.ndjson files directly from html wikipedia dumps. via the process-root command.

Note: there are 3 publicly exposed commands via typer, process-root, process-folder, process-file

process-root is probably what you want to use. It takes in the following folder structure:

dumps <- Input folder for [process-root]
 |-afwiki-NS0-20240420-ENTERPRISE-HTML <- Input folder for [process-folder]
    |-afwiki_namespace_0_0.ndjson <- Input file for [process-file]
    |-afwiki_namespace_0_1.ndjson
    |-afwiki_namespace_0_2.ndjson
    ...
 |-arwiki-NS0-20240420-ENTERPRISE-HTML
    |-arwiki_namespace_0_0.ndjson
    |-arwiki_namespace_0_1.ndjson
    |-arwiki_namespace_0_2.ndjson
    ...
 ... And so on...

Downloading and filtering the files is relatively easy.

  1. Get a list of http urls (Whichever way you prefer)
  2. Download said list (wget, curl, aria2c, etc)
  3. Extract tar files into their own folder as shown above
  4. Run process-root command.
  5. Patience.
  6. ???
  7. Finished!

wikipedia_template.py

This file contains templates used in Wikipedia articles.

If you do need to update a template, follow these steps:

  1. Open the file in your web browser.
  2. Paste the following URL, replacing <ID> with the relevant Wikidata entry ID.
https://www.wikidata.org/w/api.php?action=wbgetentities&ids=<ID>&format=json&props=labels

As for the related templates:

  • Stubs: Q4663261
  • Citation needed: Q7106262
  • Redirect: Q6042392

Note: For Sections, there are currently no templates available. These must be added manually.

mediawiki_soup.py

Before the introduction of Hugging Face's Datatrove and for the sake of simpler code development, this module implemented the MediaWikiSoup class.

This class processes HTML content into markdown format and performs additional post-processing steps on the resulting markdown.

MediaWikiSoup leverages a "filter chain" architecture. You can extend its functionalities by adding filter functions using either add_markdown_filter (for markdown processing) or add_soup_filter (for BeautifulSoup processing).

html2markdown.py

Contains a customized markdownify instance. Since this is mainly carried over from 1.5, details on it are a bit hazy.

For <a> elements, I only use the text contained. That is to say, I don't include the href.

<a href="//example.com">This is an example</a>

Will be md'd into:

This is an example

For image elements:

<img src="//example.com" alt="Alt Text"/>

Will be md'd into:

Alt Text

For <li> elements, I'm unsure what was the reason behind it. Now, God/LLM/Model/??? only knows.

folders2jsonl.py

Is a simple script converting chunked ndjson files into 1 singular file for ease of processing.

Tools

Extra tools unrelated to main filtering. But used in some shape or form.

tools/wikipedia_eligablewiki.py

As the title says, it unbiasedly selects groups of wikipedia with high enough content. Refer to Selection of Wikipedia for how it was computed.

the stats .json file can be fetched from the source page here: https://commons.wikimedia.org/w/index.php?title=Data:Wikipedia_statistics/data.tab&action=edit

Copy the json in the source into a .json file and you should be good to go.

if you have to filter a list of URLS, it should look like this:

Using mirror.accum.se as mirror:

https://mirror.accum.se/mirror/wikimedia.org/other/enterprise_html/runs/20240420/amiwiki-NS0-20240420-ENTERPRISE-HTML.json.tar.gz
https://mirror.accum.se/mirror/wikimedia.org/other/enterprise_html/runs/20240420/amwiki-NS0-20240420-ENTERPRISE-HTML.json.tar.gz
https://mirror.accum.se/mirror/wikimedia.org/other/enterprise_html/runs/20240420/angwiki-NS0-20240420-ENTERPRISE-HTML.json.tar.gz
https://mirror.accum.se/mirror/wikimedia.org/other/enterprise_html/runs/20240420/anwiki-NS0-20240420-ENTERPRISE-HTML.json.tar.gz

Or with the official dumps:

https://dumps.wikimedia.org/other/enterprise_html/runs/20240420/amiwiki-NS0-20240420-ENTERPRISE-HTML.json.tar.gz
https://dumps.wikimedia.org/other/enterprise_html/runs/20240420/amwiki-NS0-20240420-ENTERPRISE-HTML.json.tar.gz
https://dumps.wikimedia.org/other/enterprise_html/runs/20240420/angwiki-NS0-20240420-ENTERPRISE-HTML.json.tar.gz
https://dumps.wikimedia.org/other/enterprise_html/runs/20240420/anwiki-NS0-20240420-ENTERPRISE-HTML.json.tar.gz

tools/wikipedia_pageview.py

Not used in NEXT, but included. The idea is to accumulate all pageviews and filter each article based on pageviews. While it's a neat idea, I just didn't use it.

tools/wikipedia_mediaalias.py

Pretty sure it's unfinished. I didn't use it in the end. Similar to pageview. Though someone could improvise and use it.