Using Fine-Tuned LLMs to Identify Indicators of Vulnerability in UK Police Incident Logs
Researchers tested if fine-tuned Large Language Models (LLMs) could estimate the prevalence of mental ill health, substance misuse, alcohol dependence, and homelessness from unstructured UK police incident narratives. The study used nearly 3,000 logs to create estimates like 8% for alcohol dependence. While LLMs can provide scalable estimates for policy planning, their individual predictions are imperfect and require human oversight.