Talk:Machine Learning

LiftWing for Rater

Hi! What challenge should I replace the following path with?
https://ores.wikimedia.org/v3/scores/ruwiki?models=articlequality&revids=138105762
Source: https://github.com/In1quity/rater/blob/master/rater-src/api.js#L18 Iniquity (talk) 08:33, 23 December 2025 (UTC)

A suggestion for a model

I have an idea for a tool that can use an LLM to help with editing a Wikimedia wiki. Is there some LLM that I should use for this? Its features should be:

1. As close as possible to Free Software (what it means with LLMs is kind of fuzzy, but I want to at least try to respect Wikimedia's Free Software commitment).
2. Should be able to reasonably support English and, let's say, the top 30 languages on https://meta.wikimedia.org/wiki/List_of_Wikipedias (if it supports more, it's good, of course). I don't expect any LLM to support any language perfectly, but it should be able to do some basic grammar checking in any of those languages, and translation between any of those two languages, and those checks or translations wouldn't be totally useless nonsense for the human who will examine them.
3. It must be able to handle hundreds of API requests per day (if it can handle thousands or more, it's better).
4. It must be either publicly available online or runnable on my own server (for example in Toolforge).

Perhaps the WMF or some other Wikimedia affiliate already hosts such a thing? For example, maybe the one that was made for Extension:ArticleSummaries is still alive somewhere? (The idea I have is completely different from ArticleSummaries, but the backend can be the same.) Amir E. Aharoni {{🌎🌍🌏}} 15:18, 23 February 2026 (UTC)

Interested in ML for citation and content quality

Hi, I'm a relatively new contributor with a software background. I want to share where I started, where I got pushback, and where I've ended up, because the journey taught me a lot.

I first built an AI tool that did a full review of Articles for Creation drafts: a source-by-source notability assessment, neutrality and conflict-of-interest flags, and an overall AfC-readiness verdict with the governing policy cited for each finding (github.com/orincolor/wikiafcreviewer). When I floated it with the editing community, I got clear and fair pushback about using LLMs to make editorial judgments, and I took it on board.

So I narrowed it down a lot. The tool I work on now, CiteCheck, does one small and transparent thing: it flags whether a citation actually exists and matches its source, to catch fabricated or mismatched references. No notability calls, no accept or decline, no editorial judgment. Each flag shows its evidence, and a human verifies and decides.

In parallel I've been doing the actual work by hand: verifying and fixing citations on existing articles, to learn the policies and culture from the inside. That was the advice I was given and it has been genuinely useful, including learning how often an automated flag is wrong until a human checks it. I have caught and fixed several false-positive patterns in my own tool exactly that way.

I'd really like to learn from this team, since you've built ML tools the community actually trusts (ORES and Lift Wing, the revert-risk models, Automoderator). Two questions whenever someone has a moment:

1. How did those tools earn community acceptance? I'm trying to understand the consultation and trust-building path, not the engineering. 2. Is there a useful way for someone like me to start contributing in the citation, verifiability, or quality-tooling space? Happy to begin small.

Thanks, and no rush. Ordiopside (talk) 17:51, 12 June 2026 (UTC)

@Ordiopside Hello, thank you for your interest in Wikipedia. I believe you may be interested in Edit check, which is developed by the Wikimedia Foundation, and en:User:Alaexis/AI Source Verification, a tool developed by an experienced editor. (Disclaimer: I am not the staff from Wikimedia Foundation and can't represent them) SCP-2000 (talk) 13:37, 17 June 2026 (UTC)

@Ordiopside Come post over on the talk page at Wikipedia:WikiProject AI Tools! Several people there are interested in using LLMs as part of tools that editors can apply to make tasks more efficient or fun. Dreamyshade (talk) 06:28, 30 June 2026 (UTC)

AI assisted image background clean up

CC Aiko Chou, Kevin Bazira, Georgios Kyziridis, Tobias Klausmann, but feel free to ping other relevant folks.

Hello ML team, allow me to share with you the following discussion : c:Commons_talk:CropTool#AI_assisted_background_clean_up. It's just to share with you another usecase / wishlist.

• 
  Step 1: photography the produce on white page, light must come from behind the photographer.
• 
  Step 2: image upload and prompt request.
• 
  Step 3: elegant Image is ready. Step 4: Download, crop, upload to Commons.

Thanks for the good push ! Yug (talk) 08:01, 12 July 2026 (UTC)

Question about documentation

I'm looking at wikitech:Machine Learning/LiftWing#Current Inference Services and meta:Machine learning models#Production models to learn about models that I could use for article quality analysis tools for WikiProjects, and I'm trying to figure out if these lists should be synchronized, or if they're different on purpose? For example, the list on Wikitech doesn't have meta:Machine learning models/Production/Multilingual reference need and meta:Machine learning models/Production/Multilingual revert risk. Dreamyshade (talk) 17:49, 23 July 2026 (UTC)