Wikidata:Quonter Vandal

Quonter Vandal (pronounced "Counter Vandal") is a tool for identifying vandalous edits to Wikidata.

You can find the tool here: Quonter Vandal.

This is a tool written by BrokenSegue (talk). Please direct comments to the talk page.

Notes

• Currently running using ChatGPT (Q115564437) but this will likely change to an open source / free model
• Trained on ~500 hand labeled examples and roughly 95% accuracy over said corpus
• Works for any kind of diff to items (not just changes to label / description / alias). Considers groups of edits on an item by a user at one time.
• Waits a few minutes after the edit to make sure the user is done editing before reporting the edit as vandalous
• Currently BrokenSegue is paying the hosting fees for this service

FAQ

• Where are the "rationales" drawn from?
  • The rationales are generated by the model itself. They may be wrong or hallucinated.

• Which edits does this look at?
  • Only edits by anons / unverified users

• The model seems to get confused by edits like XXX?
  • Yup having only been trained on ~500 examples there's lots of kinds of edits it doesn't know how to handle

• The site shows me edits in languages I don't know?
  • Yeah the model is multi-lingual. I personally use google translate to patrol in other languages (being cautious when it could be a mistranslation). In the future filtering to language could be an option.

• The website is janky in various ways?
  • Yeah the website is just notional. It would need to be a lot nicer (I want to rewrite it entirely) but this is just a proof of concept to see if there's interest. Please do report bugs in the website though.

• Is this open source?
  • Yes see, the github.