We will evaluate our content policy models at flagging edits that fall into the category of silly vandalism, as described on https://en.wikipedia.org/wiki/Wikipedia:Vandalism.
Build a pipeline for fetching edits from certain date ranges, classifying them, and calculating accuracy metrics.
Recent edits pipeline + calling other classifiers added in this commit, and subsequent commits: https://gitlab.wikimedia.org/repos/product-safety-and-integrity/data-analysis/-/commit/41dbdbf83ede7caf00818213a36a81c6c8edcb32
All wikis were tested. Next we will do community labelling of some of the edits on enwiki and iterate on the prompt, then re-rest on all 5 wikis.