Code and data for Optimizer and residual settings rather than label fidelity drove low fidelity pretraining effects in a fractional physics informed network benchmark

Abstract

Reproducibility archive accompanying the revised manuscript. Contains PyTorch code, validated reference solutions, low-fidelity labels, frozen experimental protocols, run configurations, trained models, selected checkpoints, predictions, analysis scripts and numerical audits. Includes 296 models from claim-check studies CC1–CC4. The study examines low-fidelity pretraining in a one-dimensional fractional diffusion–reaction benchmark at derivative orders 0.5 and 1. Apparent benefits and harms depended strongly on pretraining optimizer and residual settings. Under the combined tested corrections, pretraining did not outperform physics-only training. Replacing low-fidelity labels with reference values produced little change at the tested fidelity levels. Conclusions are limited to this benchmark and these configurations. Download all 71 ZIP files and extract them into the same directory. README.txt provides instructions, SHA256SUMS verifies downloads, and FILE_MANIFEST.json lists the archived contents.