Using Instruction-Tuned Large Language Models to Identify Indicators of Vulnerability in Police Incident Narratives