Page MenuHomePhabricator

Update Revise Tone pipelines to incorporate additional languages and update topic filters
Closed, ResolvedPublic

Description

Part 1: Removing topic filters

Let's remove all the topic filters that are currently in place. We can do this gradually if needed, in order to scale our system adequately to handle the increase in requests.

Part 2: Scaling to more languages

For the initial launch of the Revise Tone experiment, we only generated tasks for the wikis on which the experiment would run (en, fr, ar, es). Now we're ready to expand the Revise Tone task across all languages where the model is supported.

To do this, we need to enable the additional wikis in 2 places in the pipeline: the generation step, and the invalidation step.

As Aiko determined here, we don't need to manually exclude the sections of each of these wikis that should not be included in suggestion generation, because this was only needed when the pipeline was based on wikitext.

The Growth team has also indicated that we don't need to pre-generate a dataset of suggestions that can be ingested and made available when we turn on Revise Tone in the new wikis. Instead, suggestions will be created more gradually, as articles are edited.

Event Timeline

Sucheta-Salgaonkar-WMF renamed this task from Update Revise Tone generation and invalidation pipelines to incorporate additional languages to Update Revise Tone pipelines to incorporate additional languages and update topic filters.Aug 3 2026, 3:01 PM
Sucheta-Salgaonkar-WMF updated the task description. (Show Details)
Sucheta-Salgaonkar-WMF updated the task description. (Show Details)

Change #1324586 had a related patch set uploaded (by Bartosz Wójtowicz; author: Bartosz Wójtowicz):

[machinelearning/liftwing/inference-services@main] revise-tone-task-generator: Update allowed topics.

https://gerrit.wikimedia.org/r/1324586

Change #1324586 merged by jenkins-bot:

[machinelearning/liftwing/inference-services@main] revise-tone-task-generator: Update allowed topics.

https://gerrit.wikimedia.org/r/1324586

Change #1324702 had a related patch set uploaded (by Bartosz Wójtowicz; author: Bartosz Wójtowicz):

[operations/deployment-charts@master] ml-services: Image bump for revise-tone-task-generator.

https://gerrit.wikimedia.org/r/1324702

Change #1324702 merged by jenkins-bot:

[operations/deployment-charts@master] ml-services: Image bump for revise-tone-task-generator.

https://gerrit.wikimedia.org/r/1324702

We've done a first bump of allowed topics to 30 out of 64 topics covering Culture, History_and_Society topics plus STEM.Technology and there are no problems so far.
We'll let it run for a day to make sure everything is working smoothly and we'll enable remaining topics if there will be no problems.

Change #1325289 had a related patch set uploaded (by Bartosz Wójtowicz; author: Bartosz Wójtowicz):

[machinelearning/liftwing/inference-services@main] revise-tone-task-generator: Remove topic filtering.

https://gerrit.wikimedia.org/r/1325289

Change #1325289 merged by jenkins-bot:

[machinelearning/liftwing/inference-services@main] revise-tone-task-generator: Remove topic filtering.

https://gerrit.wikimedia.org/r/1325289

Change #1325423 had a related patch set uploaded (by Bartosz Wójtowicz; author: Bartosz Wójtowicz):

[operations/deployment-charts@master] ml-services: Update revise-tone-task-generator.

https://gerrit.wikimedia.org/r/1325423

Change #1325423 merged by jenkins-bot:

[operations/deployment-charts@master] ml-services: Update revise-tone-task-generator.

https://gerrit.wikimedia.org/r/1325423

We have now disabled the topic filtering alltogether.

The latency seems to have neither increased nor decreased - the fact that we don't do topic predictions is being balanced by doing more html page fetches for the new topics. We'll keep monitoring the performance of the model and the number of new tone recommendations being sent. If everything goes smooth, we will enable the additional languages from part 2, probably in the next week.

Our tracking of the task pool size seems to be showing a change in the trend 😊

image.png (745×305 px, 20 KB)

Regarding Part 2: Scaling to more languages, the Growth team would like to enable the following additional languages, according to T426364:

Languagewiki_dblang_code
Russianruwikiru
Japanesejawikija
Dutchnlwikinl
Spanisheswikies
Chinesezhwikizh
Germandewikide

Change #1327514 had a related patch set uploaded (by Bartosz Wójtowicz; author: Bartosz Wójtowicz):

[machinelearning/liftwing/inference-services@main] revise-tone-task-generator: Enable additional languages.

https://gerrit.wikimedia.org/r/1327514

Change #1327514 merged by jenkins-bot:

[machinelearning/liftwing/inference-services@main] revise-tone-task-generator: Enable additional languages.

https://gerrit.wikimedia.org/r/1327514

Change #1328401 had a related patch set uploaded (by Bartosz Wójtowicz; author: Bartosz Wójtowicz):

[operations/deployment-charts@master] ml-services: Update revise-tone-task-generator.

https://gerrit.wikimedia.org/r/1328401

Change #1328401 merged by jenkins-bot:

[operations/deployment-charts@master] ml-services: Update revise-tone-task-generator.

https://gerrit.wikimedia.org/r/1328401

Change #1328405 had a related patch set uploaded (by Bartosz Wójtowicz; author: Bartosz Wójtowicz):

[operations/deployment-charts@master] changeprop: enable additional languages for revise-tone-task-generator

https://gerrit.wikimedia.org/r/1328405

Change #1328405 merged by jenkins-bot:

[operations/deployment-charts@master] changeprop: enable additional languages for revise-tone-task-generator

https://gerrit.wikimedia.org/r/1328405

We've just enabled the new languages (es, ru, ja, nl, zh, de) for the Revise Tone pipeline.
Service logs confirm we're successfully processing edits from all new wikis. We will keep monitoring service metrics, but everything looks healthy so far. I am resolving the ticket :)