MLSD — Harmful / Weapon-Sales Content Detection

Integrity-style MLSD: detect harmful content (weapon sales, harmful ads) at scale. Weak labels, active learning, policy alignment, and precision-recall trade-offs are the live discussion points. Requirements Functional: classify posts / ads as harmful with sufficient precision to action; minimize false-positives that violate creators. Pipeline: text + image multi-modal embedding → binary or multi-label classifier; downstream policy mapping. Labels: small high-precision human-reviewed set + large weakly-labeled set (heuristics, user reports); active-learning loop. Metrics: precision @ fixed recall (or AUC-PR); per-policy slice precision; reviewer agreement. Operational: human-review queue prioritization, model-confidence routing, feedback ingestion. Coverage of harmful-ad sub-variant: also pull advertiser signal (history, payment, account age). Notes Reports mention interviewers ask about handling label noise, policy nuance, and reviewer-disagreement resolution — not just model architecture. Be ready to argue per-policy thresholds (different precision targets for hate-speech vs spam). Preparation Memorize a weak-supervision pipeline: programmatic labels → noise-aware loss → small clean set for calibration. Prep an active-learning loop pitch: uncertainty sampling, diversity sampling, periodic reviewer assignment. Rehearse a 3-sentence answer for "how do you measure policy drift?".