How we built CoPE (blog.zentropi.ai)

A new paper detailing the development of CoPE, a 9-billion parameter model, introduces an innovative training technique aimed at enhancing content moderation without the need for retraining when policies change. CoPE achieves approximately 91% F1 score on hate speech detection, outperforming the larger GPT-4o model, while operating at just 1% of its size and achieving low latency. This is significant as traditional content classification systems struggle to keep up with evolving policy requirements, often causing enforcement delays. The core of CoPE's methodology lies in "Contradictory Example Training," where the model learns to evaluate content against varying policies that yield opposite labels for the same input. This approach enables the model to interpret complex policy nuances, rather than relying solely on pattern matching. Additionally, the authors employed a novel dataset creation method called "binocular labeling," which streamlines the labeling process by generating multiple policy versions and focusing manual efforts only on ambiguous cases. The authors highlight ongoing challenges in policy evaluation and multilingual application, opening avenues for collaboration in refining these methods within the AI/ML community.