A learnable detector-space metric for Timepix3 clusters: code, data and results

Companion deposit for the JINST article "A learnable detector-space metric for Timepix3 clusters: pixel-level particle identification and what the learned weights do and do not measure".Contains the cluster-level Timepix3 measurements used for training and for the blind test, the complete training and analysis pipeline, all 65 trained checkpoints and every per-run summary behind the published tables and figures. Both passes of the six-configuration ablation are included: the one in which the absolute pixel timestamps were parsed in single precision — which leaves the in-cluster time-of-arrival identically zero for 98.6% of the clusters — and the corrected one in double precision, from which the published numbers are taken. The comparison between the two is used in the paper as a measurement of the run-to-run reproducibility of the pipeline (0.040 in macro F1), and is the tolerance on any single number quoted from a training.Data: 14,716 clusters from Po-210 (alpha), Sr-90 and Cs-137 (e-) and Fe-55 (gamma) sources acquired with two Timepix3 detectors, plus 3,080 Am-Be clusters used only for a blind consistency test. Time-over-Threshold is uncalibrated and no time-walk correction was applied. Reproducing the tree baselines to the digit requires Python 3.12.4, numpy 1.26.4 and scikit-learn 1.4.2; see README.md and REPRODUCE.md.