System Design · Meta · Hard
Requirements Functional goal: identify harmful posts and advertisements accurately enough to support enforcement, while keeping creator-impacting false positives as low as possible. Pipeline: generate joint text-and-image embeddings, feed them into a binary or multi-label model, and then apply downstream policy decisions. Labels: combine a small, carefully human-validated dataset with a much larger weak-label dataset built from signals such as heuristics and user reports;…
Checking your access…