Making Sense of Sensors: Improving LLM Interpretation of Time-Series Data

The repository contains the datasets and evaluation outputs used in this study. synthetic_timeseries Synthetic water tank distance signals used to generate the evaluation dataset. The signals are based on patterns observed in real deployments and represent different water consumption behaviors, including stable, cyclic, and irregular usage patterns. llm_responses Responses generated by the language models under each prompt configuration (BB, FB, FA) for the 24 evaluation segments. chatgpt_5_2_evaluations Evaluation results produced by ChatGPT-5.2 acting as an independent evaluator model. Each file contains rubric scores and failure-mode flags for all segments. gemini_3_evaluations Evaluation results produced by Gemini-3, including dimension-level scores and failure-mode flags. sonnet_4_5_evaluations Evaluation results produced by Sonnet-4.5, providing independent scoring of responses across all evaluation dimensions.