Expert-Level Crisis Detection in Mental Health Conversations

In three linesCRADLE-Dialogue, a clinician-annotated benchmark of 600 dialogues, evaluates turn-level crisis detection in conversations. The dataset distinguishes early warning signals (Alert) from explicit confirmations (Confirm) for suicide ideation, self-harm, and child abuse. A 32B model outperforms existing open-source models with Micro F1 scores of 40-60%.

Summary generated by Claude — human-verified