Mitigating LLM Policy-Induced Bias in Sexual Orientation Identification Using Cue-Augmented Fine-Tuning

Large Language Models (LLMs) can often exhibit policy-induced bias when classifying sexual orientation in clinical texts, leading to fairness and safety concerns. To address limitations introduced by general-purpose safety alignment and domain mismatch in LLMs when applied to clinically sensitive attributes, we fine-tuned LLaMA-3.1 8B Instruct on a curated dataset of sexual orientation cues using LoRA and prompt-based supervision. Our method improved classification performance and reduced false associations. At the note-level, our model achieved 99.17% accuracy (macro F1=0.9923) across three classes (straight/men who have sex with men/uncertain). At the patient-level, binary classification reached 98.48% accuracy (F1=0.9848). These results demonstrate that targeted fine-tuning can significantly mitigate policy-induced bias in LLMs sexual orientation classification for downstream clinical applications.