When we released CoPE-B last week-- our newest language model for content classification-- the standard metrics showed only modest gains. But every time we used the model it *felt* like a huge step up from our prior model, responding to policy subtlety and changes with much greater fidelity.
