TADA toy sharpen experiment — HDF5 training data (ALASKA source + sharpen target, QF100, UERD 1 bpnzac)

HDF5 datasets for reproducing the toy sharpen experiment of: Abecidan, R., Itier, V., Boulanger, J., Bas, P., & Pevný, T. Tackle CSM in JPEG Steganalysis with Data Adaptation. arXiv:2605.21523 (2026). Companion code: https://github.com/RonyAbecidan/TADA Contents--------1. color_raws_512.hdf5 - Key: train - 2,000 color TIF crops (512×512) from ALASKA RAW (amaze demosaicking), selected to be as spatially uniform as possible (raw_uniform pool). - Used as the TADA source side 2. targets/sharpen_full_stego.hdf5 - Keys: operational, pmap_ope, eval, pmap_eval - Toy sharpen target: grayscale, 3×3 sharpen kernel, JPEG QF100, full UERD embedding at 1 bit per non-zero AC DCT coefficient (bpnzac). - 1,000 operational images (+ pmap_ope) for TADA training; 1,000 eval images (+ pmap_eval) held out for disjoint evaluation. Usage-----Place the files as documented in the TADA repository and run: python pipeline_learning_config.py sharpen100_bpnzac.yaml cuda Provenance----------Built from ALASKA#2 RAW images. ALASKA dataset: Cogranne, Giboulot, Bas — The ALASKA Steganalysis Challenge (IH&MMSec 2019). Notes------ These files are not required for the operational YFCC100M experiments in the paper.- Image rights follow ALASKA#2 terms; this deposit is for research reproducibility only.