Interpreting urban parking regulations from natural language text is a nuanced reasoning task, requiring precise understanding of temporal conditions, day-of-week constraints, and exceptions. This paper presents an improved pipeline that combines advanced traditional machine learning methods with fine-tuned modern large language models (LLMs) for classification of parking rules. This work creates a dataset from the New York City Department of Transportation’s Parking Regulation Locations and Signs database, which is updated daily. After filtering for relevant rules, merging multiple signs per location, and cleaning metadata, obtain 64,576 entries were obtained. From this, a balanced subset of 1,674 examples is created, each paired with a fictitious “current time” prompt and binary label, generated by GPT-4 and partially verified by human inspection to ensure handling of overlapping and conflicting cases. This work evaluates both traditional models and a fine-tuned Meta-LLaMA-3 model using LoRA adapters and 4-bit quantization. Results show that XGBoost achieves 91% accuracy and 0.94 AUPRC, outperforming all other baselines, while the fine-tuned LLaMA-3 reaches 94% accuracy and excels at handling semantically complex, temporally bound rules. Incremental data experiments reveal that traditional models benefit strongly from larger training sets, with XGBoost approaching LLM performance beyond 1,300 examples. This work extends prior studies by improving baseline selection, expanding interpretability analysis, and systematically comparing scaling behavior between classical and LLM-based approaches. Potential applications include deployment in mobile or camera-assisted parking guidance systems.