VideoRAPL: A Dataset and Benchmark for Video Workload Inference from Intel RAPL Power Traces

VideoRAPL is a privacy-reviewed dataset and benchmark for inferring video workloads from Intel RAPL processor-domain power traces. Version 1.0.0 contains 589 sanitized traces (472 controlled FFmpeg and 117 manual real-world captures), 1,142,410 sample rows, 40,163 model windows, eight canonical trace-level benchmark split sets, one exploratory split, reproducibility code, reference results, complete experiment artifacts, figures, and model weights. The deposit is divided into a core archive and an optional artifact overlay that extract into the same VideoRAPL_v1.0 tree. Source videos are not redistributed. Exact wall-clock timestamps, private run identifiers, personal paths, and the private source-to-public mapping are excluded. Data, documentation, results, figures, and model artifacts are licensed under CC BY 4.0; packaged code is licensed under MIT. Documentation and source code are available at https://github.com/Abdelhaiben/VideoRAPL.