Page MenuHomePhabricator

Retrain damaging/goodfaith models for huwiki
Closed, ResolvedPublic

Description

Train a new model using the new labeling campaign for huwiki. https://github.com/wikimedia/editquality/pull/195/files

Event Timeline

Halfak created this task.Jul 15 2019, 4:38 PM
Restricted Application added a project: artificial-intelligence. · View Herald TranscriptJul 15 2019, 4:38 PM
Restricted Application added a subscriber: Aklapper. · View Herald Transcript

@Halfak the models have been retrained, let me know your thoughts about the tuning reports
https://github.com/wikimedia/editquality/pull/207

I just reviewed. Just a minor note on file naming. The fitness statistics look good to me. I think we're running into issues with the fragility of our AUC measures. If you look at the ROC AUC and the PR AUC of the target classes (true for damaging, false for badfaith) then it looks like we're getting a nice improvement here.

Halfak closed this task as Resolved.Aug 7 2019, 3:53 PM
Samat awarded a token.Aug 7 2019, 5:47 PM