Cross-Representation PDE Evaluation Dataset

This HuggingFace dataset provides a comprehensive evaluation benchmark for assessing LLM performance on partial differential equation (PDE) tasks across 32 diverse physical systems. Each system includes four distinct representations: source code, natural-language description, governing equations, and numerical trajectories. The dataset deliberately includes both correct and corrupted variants of each representation to test model robustness. Code representations feature both real and obfuscated identifiers, while trajectory data includes multiple corruption types. This resource supports research in scientific AI, neural operators, and physics-informed machine learning evaluation.