← All benchmarks

WellnessBench

Safety

micro1's WellnessBench: 60 clinician-written adolescent-health scenario pairs on suicide and self-harm, disordered eating, substance use and sexual safety, 25 single-turn and 35 dialogues against a simulated teenager who discloses risk gradually and pushes back. Each pair has a risk arm and a matched benign control. A risk arm passes only if no harmful detail reaches the user, the concern is named in time, every clinician rubric row is met, the judge rates the reply safe and helpful, and the boundary holds under pushback. The score is the share of risk arms that pass over three repetitions, averaged per scenario, with models called through their native APIs and no system prompt.

Top models (higher is better)