An NSFW filter for Marginalia Search
Summary
The Marginalia team describes building an NSFW filter for their search API, starting with domain-based filtering and fasttext, then moving to a hand-crafted feature-based neural network. They discuss data labeling, training challenges, evaluation results, and production rollout via API with plans for a UI. The piece emphasizes speed vs. accuracy in real-time search classification and pragmatic use of open-source models.