The Internet of Things generates massive heterogeneous data in smart cities, energy management, and environmental monitoring, creating an urgent need for efficient, scalable data standardization. Accurate classification of smart data models (SDMs) is essential for interoperability, yet existing methods often incur high computational costs that exceed the capabilities of edge devices.
This study evaluates a range of lightweight open‑source language models for mapping an input data entity to its best‑fitting SDM under resource‑constrained conditions. The benchmark covers general‑purpose (GP), reasoning‑specialized (RS), and code‑specialized (CS) architectures across several domain‑specific datasets. Results show that certain small models retain acceptable accuracy while dramatically reducing memory usage and inference latency, making them suitable for edge deployment.
To address the literature gap regarding low‑cost solutions, the work compares these models against two near‑zero‑cost similarity baselines—TF‑IDF and a lightweight sentence encoder. The baselines provide reference performance with minimal resources, helping to quantify the practical value of LLM‑based classification on edge platforms.
Overall findings suggest that model selection should consider task characteristics, resource budget, and desired accuracy; the task is best framed as entity‑to‑model matching rather than generative QA; and deployment can benefit from techniques such as model distillation or quantization to further shrink footprint.
Review: The paper delivers a systematic evaluation framework and actionable guidance for deploying small language models on edge devices to classify smart data models, filling an important gap in current research.