Hive Vision Language Model

Multimodal model that enforces a policy written as a prompt across images, video and text.

About Hive Vision Language Model

Hive states that Trust and Safety teams can use it to apply their own moderation rules across different content types.

Hive lists sexual and suggestive content, hate and abusive content, drugs, weapons and age-related policy enforcement among its policy areas. Yubo, an online social space for youth, is a Hive customer, and Hive states that it partnered with Yubo from the app's initial product launch four years ago. Tango, a live streaming platform, used Hive's VLM to analyze multiple frames simultaneously.

Drafted from hivemoderation.com · not yet confirmed by Hive

How it works

Instead of a fixed set of labels, Hive's Vision Language Model takes the policy to enforce as a prompt and applies it to images, video and text, alone or together, so text and images that change each other's meaning are judged as one. When a policy changes, the prompt is edited and the model is not retrained. Hive lists sexual content, hate, drugs, graphic content and weapons among the areas it covers.

Video can be checked frame by frame or with one prompt across the whole clip. Hive offers a Playground for testing prompts and AutoML Evaluations for structured testing, and the model is available for production use without an enterprise contract.