DrowsyFormer-Gen — revision code and run records

Abstract

Pipeline code and aggregate run records for the revised manuscript "DrowsyFormer-Gen: A Triple-Stream Transformer with Diffusion Augmentation for Driver Drowsiness Detection, Evaluated for Recall, Calibration and Pre-Onset Alarm Coverage", submitted to Scientific Reports. Every number and every figure in the manuscript is produced by the code here from the stored run records: analysis.py resolves the tables from results/runs/*.json, and submission_figures.py writes the eight submitted figures as Figure1-Figure8. The paper has 20 display items (Figures 1-8, Tables 1-12); the former Supplementary Information has been merged into the main text. Contents- code/ — 12 modules: preprocessing and subject-grouped windowing, the triple-stream model, the class-conditional 1-D DDPM, classical baselines, metrics and calibration, event-coverage analysis, and the figure scripts- results/runs/ — 25 run records, one per (dataset, method or ablation variant, seed)- results/ — resolved tables (report.json), corpus and importance analyses, diffusion chain traces- figures/ — Figure1-Figure8 as submitted, plus the working set they are selected from- SHA256SUMS.txt, requirements.txt Not included- Participant data. The DMPS and SC-DD workbooks are special-category human-subject records collected under institutional protocol VT/ERC/2022-104. They are not redistributed here and are available from the corresponding author on reasonable request, subject to institutional approval and a data-use agreement.- Fold checkpoints and per-window neural predictions from the reported runs were not retained. Aggregate per-fold metrics are in results/runs/. Supersedes 10.5281/zenodo.20488319, which contained only a synthetic-sample generator, four schema-matched example workbooks and a placeholder where the notebook was meant to be. That record cannot reproduce the results and should not be used.