|Objective Revision Evaluation Service (ORES)|
ORES supports a limited set of Wikis, but support is growing all of the time. This page describes how to get ORES support for your favorite wiki. Make sure to double-check the support table as we may already support your wiki. There are several different types of support that we can provide. It really comes down to which class of model and the tier of support we are aiming for. If you need help figuring out how to make a request, post on the talk page.
In order to set up basic support for your wiki, the first thing we'll need to do is set up the language assets that ORES will need to operate in your wiki. Check if your language is listed in our code or check phabricator for open tasks. If you don't find your language listed, use the link below to request support!
At the very basic level, we provide a
reverted model that attempts to predict whether or not an edit will need to be reverted. This model is "trained" using a sample of past reverted edits that happened in a particular wiki. Once we have basic language support ready, we'll set up a
reverted model for your wiki.
reverted model is useful and can be trained based on the history of article, it is slightly problematic. It's much better if we can train our prediction models on more nuanced judgements of the quality of an edit. Many of the tools that use ORES to support reviewing recent changes require this level of support. The
damaging model predicts whether an edit causes damage and the
good faith model predicts whether an edit was saved with good intentions.
To gather the data needed to train these models, we can set up a Wiki labels campaign with a random sample of edits for evaluation. See the ongoing campaigns to join one (preferably if you have editing experience), or request a labeling campaign below.
We train article quality models based on quality assessments (like en:WP:1.0 in English Wikipedia) if they are available. If not, we'll need to set up a Wiki labels campaign and have editors assess a sample of articles (usually ~5000) in order to train ORES how to make assessments.
We train draft quality models based on deletion reasons that are recorded in the deletion log. In order to start work on the model, we'll need to ask for some help identifying what types comments are present in deletion reasons that we might flag as deeply problematic.