Repeated-measurement-aware reliable surface roughness prediction dataset and code

This dataset and code release supports the manuscript ‘Repeated-measurement-aware reliable prediction of image-based surface roughness’. It contains 330 physical metallic samples from aluminium, titanium alloy, stainless steel 316 and brass; 3,300 surface images (10 local views per sample); and 1,650 profilometer records comprising five repeated measurements of Ra, Rv, Rz, Rt and Pt per sample. The release includes processed targets and repeat-variation descriptors, image-to-sample manifests, fixed five-fold and low-data splits, leave-one-material-out splits, model source code, experiment configurations, reported predictions and summary tables, and scripts for validation and reproduction. The regression target is the middle-three trimmed mean of five readings. The sample standard deviation of all five readings is provided as an empirical repeat-variation descriptor rather than a complete metrological uncertainty estimate. Data and images are licensed under CC BY 4.0; source code is licensed under the MIT License.