PVminerLLM2: Improving Structured Extraction of Patient Voice via Preference Optimization

PVminerLLM2 introduces an enhanced framework for structured extraction of patient-generated text, utilizing preference optimization to mitigate token-critical errors that traditional supervised fine-tuning struggles with. Key innovations include a token-level gated stabilization term and confusion-aware preference pair construction, along with token-importance weighting to address class skew. Benchmark results show PVminerLLM2 surpassing previous models by up to 4.43% in various extraction tasks, highlighting its potential for improving patient-centered outcomes research in AI applications.