A Lightweight Large Language Model for Personal Sleep Quality Estimation

This paper presents a lightweight large language model (LLM) framework for personalized health advice generation, addressing critical limitations in resource-constrained scenarios. To resolve the inherent conflict between model performance and computational costs in existing solutions, we propose a threefold innovation: First, a structured prompting template dynamically integrates multimodal health data with semantic instructions. Raw sensor data undergoes anomaly detection and temporal aggregation. This process creates concise, semantically rich inputs for the LLM and eliminates the need for manual feature engineering. Second, we enable parameter-efficient fine-tuning of LLaMA 3.1 using Low-Rank Adaptation (LoRA). This approach injects rank-16 matrices into both attention and feed-forward layers, achieving single-GPU fine-tuning with only 7.93 GB memory overhead and sub-8-minute convergence. Third, we introduce an arena-style evaluation framework to assess health advice personalization. This framework uses third-party LLMs as neutral judges to compute ELO ratings through pairwise blind comparisons of eight candidate models, providing a quantifiable metric. Experimental results demonstrate that our optimized model attains a 36% ELO score improvement (from 1,072 to 1,389), surpassing some closed-source models as well as all open-source smaller models.