Lightweight Contextual LLMs for IoT Data Interpretation in Smart Cities

While the populations in cities are growing, the need for smart cities to optimize resources and life quality is increasingly a pressing matter. At the heart of this transformation lie the IoT devices creating massive volumes of data to be interpreted in real time. This paper introduces the Lightweight Contextual Large Language Models designed for efficient processing of IoT data in Smart Cities. In the approach proposed, advanced model pruning and knowledge distillation are used, keeping the solution computationally efficient with no sacrifice in interpretative power. We validate our model's effectiveness on the UrbanSensing dataset in anomaly detection, predictive maintenance, and analytics of engagement on tasks involving citizens. Our framework allows feasibility on edge devices with a 40% reduction in inference time and up to 50% decrease in computational overhead, thus guaranteeing scalability and sustainability in real-world smart city applications. The findings bridge the gap from computational efficiency to contextual understandings, thereby offering the premise for enhanced urban resilience and wiser decision-making.