AI-Powered IoT Assistant for Electric Vehicles: Fine-Tuning a Small Language Model for Intent Recognition and Response Generation

The integration of artificial intelligence (AI) and Internet of Things (IoT) in electric vehicles (EVs) has revolutionized vehicle automation, user interaction, and energy efficiency. This paper presents an AI-powered IoT assistant designed to enhance EV user experience by understanding and responding to natural language queries. We fine-tune a Small Language Model (SLM), Phi-3.5 Mini Instruct, using Supervised Fine-Tuning (SFT) with QLoRA, optimizing it for intent recognition, entity extraction, and response generation. A synthetic multi-turn conversation dataset with 50 intents and entities was generated using GPT-4o, covering various EV-related tasks such as battery monitoring, charging station navigation, and vehicle control. The model was fine-tuned using 861 training samples, 50 validation samples, and evaluated on 100 test samples. The training process, executed on four A10G GPUs, achieved a training loss of 0.18 and validation loss of 0.21. Evaluation results demonstrated significant improvements in intent recognition, with precision, recall, and F1-score increasing from the 40s to the 70s after fine-tuning. Additionally, response quality assessment using GPT-4.0 showed notable enhancements in accuracy, coherence, and specificity. These results validate the effectiveness of lightweight domain-specific fine-tuning and highlight the potential for deploying efficient AI-driven assistants in smart vehicles. Future work will focus on expanding real-world training data, optimizing for edge deployment, integrating multimodal inputs, and enhancing security mechanisms. This research contributes to the advancement of context-aware, intelligent vehicle assistants, paving the way for seamless human-AI interaction in electric mobility.