I would interpret "algorithms" here as referring to entire separate machine learning models trained differently (if not trained on entirely separate data sets).

I would interpret "algorithms" here as referring to entire separate machine learning models trained differently (if not trained on entirely separate data sets). It would make sense to me that you would want to train separate models for identifying specific categories of abuse instead of training one Ultimate Model, because debugging focused models is going to be easier. That would line up with the massive number of parameters as well.