Once T416659: Host sentence-transformers/LaBSE model in LiftWing is resolved, the model infrastructure dependency for this task will also be ready

Here is nodejs client code that can give cosine similarity based on the above service.
Pull the above patch, then

docker compose build labse-embeddings
docker compose up labse-embeddings

Get this code,

The current link suggestion system has an algorithmic issue that I had pointed in 2025. In this document and in a slack conversation(private), I explain this in detail. Basically, I argue that the problem is not a machine learning problem but a distributional fitting problem.

It looks like issue is specific to that file?
I tried https://commons.wikimedia.org/wiki/File:Moon_elevation.stl#/media/File:Moon_elevation.stl and working as expected

slightly unrelated to this ticket - the sections suggested for translation from this example is incorrect for many titles due to issues in current section alignment algorithm. For example: Etymology is asked to translate, but it is already there in the target article.

The graphs are still empty. Is there anything else to fix or will it take time to refresh the data?

Yandex was removed from cxserver

The custom font embedding feature(also known as webfonts - https://www.mediawiki.org/wiki/Universal_Language_Selector/WebFonts ) was part of Universal Language Selector project. This was introduced more than a decade ago to support scripts when operating systems did not ship any fonts for them. The situation improved a lot in past several years. Embedding fonts in wikipedia pages has performance and network cost. So we gradually removed fonts as they got widely available. At present, in 2025, this feature is no longer recommended. It is also not actively maintained. There are a few fonts still in Universal Language Selector, but they are not embedded in wiki pages by default. Users can choose them from settings. But even then it is only available for Desktop and not for mobile. The Universal Language Selector is also undergoing a full rewrite this year as part of modernization.

The LLM evaluation could be a little bit systematic. Currently GPS-OSS is 20B param mode, which we are comparing with Aya Expanse 32B. And Aya Expanse is an older model too. The prompt need to be versioned to see which is better. I would focus on figuring out a prompt and LLM param range that gives reasonable score and keep the LLM choice flexible since LLMs are are moving target and there are many practical considering to pick one.

Marking as stalled for now.

I am abandoning the patches as they are outdated and I do not plan to keep them updated with the changes mentioned above.

This proposal is intended to support corrections for exact sentences.
...
if a user fixes a typo in one specific sentence, it gets fixed only for that instance (even if similar sentences exist).

Thanks. It works now. Tool is up https://api-explorer.toolforge.org/

Notes:

• Does not work with words in sentences since it require “fitting” that selection of word in to the original sentence and keep grammar accurate.(in localization we call this “avoiding lego messages”.) -This limitation also applies to translation memory for phrases(incomplete sentences). Chances of encountering the (exactly) same sentence again less in general prose.
• Examples
  • This is Basil(person) - ഇവൻ(male)/ഇവൾ(female) ബേസിലാണ് (the translation of basil get fused with affirmative ആണ് - aka agglutination. Same below)
  • This is basil(herb) - ഇത് തുളസിയാണ് (grammatical gender does not apply for things)
• Grammatical gender, subject verb agreement, agglutination and inflection are quite common outside the realm of analytical languages
• Word context issue is almost a solved problem in latest translation models, only present in MinT due to its limitation about contexts and anaphora resolution.(Broad context translation with self attention - in literature examples like “interest”, “bank” are often used)

If I dont use buildservice but copy the rust binary to toolforge product and run it with the following service template, service is starting successfullly

The Cloud-Services project tag is not intended to have any tasks. Please check the list on https://phabricator.wikimedia.org/project/profile/832/ and replace it with a more specific project tag to this task. Thanks!

Infobox historic site (https://www.wikidata.org/wiki/Q12857484) has templatedata in English, but no templatedata in Tamil. So cxserver has to extract template params from the template source code.

Another example. Here the infobox is not completely missing. It is shown as an indicator in source column.

Since LiftWing hosting plan is pending and under consideration by Machine-Learning-Team , I setup https://ovms.wmcloud.org/ with upstream docker image and above shellscript. Wrote an nginx proxy configuration for landing page, API key authentication and /api/v1 mounting. OpenAI api compatible APIs are available. This instance uses 8 CPU cores, 32GB RAM, 20GB Disk(models are in scratch volume). This wmcloud instance does not cancel this ticket, and for experimentation purpose.

This patch to prepare a WMF production image is outdated again as new version of Debian is released.

The model repository setup is easy with the new versions. OVMS can pull models from huggingface repos and setup all configurations. I wrote a shellscript to get all listed models in it and start the server. This script is used with the official docker images of OVMS. https://gerrit.wikimedia.org/r/c/machinelearning/liftwing/inference-services/+/1151645 will require updates based on this

Node 22 stabilizes the fetch API. It is now feature compatible with browsers fetch API. This is generally good, but it also adds more restrictions to what a valid http request can be.  The header field we are setting to pass the wikipedia domain to that wiki proxy is HOST (see the configuration). This is problematic because HOST is a forbidden header.

Just pointing out an additional detail on 'easy to translate':

For example, the "Add a Link" model needs to be trained on links that already exist within articles on a wiki, so there is a Machine Learning dependency and a chance that some wikis will be too small for the model to be properly trained.

Update: New version of Opevino and Openvino model server was released a few days ago. I updated my production bookworm based docker image patch: https://gerrit.wikimedia.org/r/c/operations/docker-images/production-images/+/1153988 to that version. This new version adds support for agentic AI(tool calling and MCP).

@Wangombe , I reviewed it and proposed some changes

I found one genuine reason why the progress value can be different. Suppose there are 10 source sections and you translate 5. The any value would be 0.5, If you modified all the sections, the mt value would be 0(just assume so) and human value would 1. When you come back a few days later, the source article can have 11 sections. Then any calculated at the time of restoration would be 0.45(5/11) . Have a look at the case #2 in the above comment. It is an example of this scenario.

Note that the is attribute based custom elements are not widely supported: https://developer.mozilla.org/en-US/docs/Web/HTML/Reference/Global_attributes/is

Some observations from initial attempt to debug:

Here is the screencast of everything working together: https://drive.google.com/file/d/1YDSvTm3ePv585-ittck2tYWH2XZ-AHLx/view
(MP4 video, 15mins, 88MB)

Updates: I successfully created a production docker image on top of WMF production wookworm image.

I am trying to build a production docker image with WMF debian bookworm+python3.11+openvino 2025.1.0+ OVMS 2025.1.0. I am referring the Ubuntu Dockerfile.

Thanks @kevinbazira and @isarantopoulos for these details. Very useful information.

Regarding the KServe API and Openvino model server:

I tried to integrate Openvino model server to liftwing. Learnings for the first iteration(see the above WIP patch):

This is a good idea.
I would suggest to use a simple client side memory for the recent selections - like localstorage and avoid server side storage(which require database, and big effort and resulting maintanance cost and tech debt).
I think it is acceptable to not have the recent selections when user switch browsers/computers.

We do have a machine translation system for kaa - Karakalpak powered by MADLAD-400 model.

The benefits of webcomponents compared to html tags is well documented. Webcomponents are html stanadard and supported by all browsers. Ofcourse, you can achieve the same features with html tags and javascript - with trade-offs. I would like to see them supported in mediawiki as they are part of web standard.

I used this documentation https://www.mediawiki.org/wiki/VisualEditor/Hooks for code search

The 45% thresholds are extreme case. It means you cannot publish a translation that has >45% machine translation. This demands heavy editing from translator. We also don't tell the translator about these percentages(T251887).
In my opinition, a simpler, consistant and predictable threshold level is better. If T is threshold for errors, make T-10 as the warning threshold, irrespective of value of T.

Please use same campaign parameter. The ArticlePlaceholder patch passes articleplaceholder as campaign param. CX Config and event source config should use the same. article_placeholder does not match.

Following discussions with the team, we've concluded that dynamic configuration is not a use case currently supported by Wikimedia. Our configurations are static, managed through updates and deployments. The
conf.yaml file facilitates this by allowing for varied configurations across different deployments, as it's separate from the codebase.

. It let me start translating, and only when I went to publish did it warn me that I would be overwriting an existing article. This seems like a step back from the behaviour of the old dashboard, where at least it would warn me before I started translating.

@jhsoby I was testing this again, and I don't see the issue now.
More importantly, the CX dashboard's new version is coming up. It will replace the current dashboard.
The following URL can be used to access new dashboard.
You can access it in https://no.wikipedia.org/w/index.php?title=Special:ContentTranslation&unified-dashboard=1&filter-type=automatic&filter-id=previous-edits&active-list=draft&from=en&to=nb#/

Seems working now. The blocking of retries for 30 mins for all consumers of API seems not an optimal approach. Also see T382376: PageViewInfo caches errors due to the request limit which can lead to denial of service

Thanks for that context. For the reasons I explained above, my advice would be to have DeepL client in your customized version and not in the upstream. Currently, our team or organization does not have a collaboration with DeepL. If this situation changes, we will definitely consider your client code. Thanks!

Hi, DeepL is a proprietary machine translation service and their API key has a price associated with - at least to use in the scale of current traffic of cxserver. For such external Machine translation services, we(WMF) usualy has a partnership. For example, to use Google API, WMF and Google has partnership agreement. For DeepL too, we will need such agreement(legal, branding, pricing).

In T387820#10621300, @Jdlrobson-WMF wrote:

Hello is it possible this ticket relates to the train blocker at T388467: TypeError: mtSuggestions.map is not a function ?