Abstract
Mental-health hotline pre-intake triage requires a worker to convert incomplete, ambiguous caller disclosure into a routing decision under ethical and time constraints. No existing dialogue dataset addresses this task. We systematically compare 36 mental-health dialogue resources, mapping their 531 label instances to 175 canonical concepts. The comparison reveals that no existing resource provides empirical labels for ecology, structural barriers, or service-system navigation, which social-work theory treats as central to intervention. We present PITA (Pre-Intake Triage Assessment), a triagedialogue dataset whose ontology, grounded in person-in-environment and trauma-informed frameworks, organizes caller state, ecological context, relational ethics, helper strategy, and disposition into independent label dimensions. PITA is constructed through a role-separated agentic synthesis pipeline. Its core mechanism, visibility-ledger-based blind risk adjudication, derives risk labels solely from caller-disclosed evidence and maintains an auditable boundary between hidden generation state and observable conversational evidence. The released corpus contains 3,000 dialogues (44,574 turns) with 63 training labels and 4 guardrails. Risk classification using only caller utterances achieves macro F-score 0.883 and quadratic weighted kappa 0.940 (RoBERTa-base), which confirms that risk labels are predictable from disclosed evidence without access to hidden generation state. A four-expert boundary analysis yields favorable quality ratings across six social-work dimensions (all means >3.2/5; 93.0% accept or minor-revise verdicts), and expert disagreement on risk and routing exposes the disciplinary boundaries that the ontology makes inspectable. PITA provides a dataset, ontology, and reproducible construction method for data-centric research on hotline triage dialogue