This comprehensive dataset supports a benchmarking study of physics-informed neural networks (PINNs) for proton dosimetry in plastic scintillators, developed for the CERN Beamline for Schools (BL4S) 2026 proposal by Team KAVACH. The dataset contains Monte Carlo simulation data from OpenTOPAS 4.2.0 modeling proton transport in EJ-200 plastic scintillator across 14 energies (70-6000 MeV). The core dataset includes 1,400 proton stopping power measurements with relativistic parameters (β, γ, momentum) and NIST PSTAR reference values, split into training (800), validation (400), and test (200) sets. A novel 280,000-event reversibility test validates Monte Carlo order-independence using reversed absorber geometries. PINN ablation studies reveal that simple smoothness constraints outperform complex multi-constraint approaches, achieving 13.3% better test MSE. Data is provided in flat CSV format for maximum compatibility. Applications include medical physics, radiation therapy verification, and scintillator detector development. The dataset enables reproducible benchmarking of machine learning approaches for radiation physics problems, bridging traditional Monte Carlo methods with modern neural network techniques.