v2 — RAFNet models and experiment assets

Assets

Extract the archives from the repository root; they contain their destination directories.

File	Contents
rafnet_v2_features.tar.gz	NCST train/validation/test cached features for six representations, labels, geometry and opaque grouping IDs
rafnet_v2_checkpoints.tar.gz	100 cached-feature checkpoints: 20 configurations × 5 seeds, plus portable configs and run metrics
rafnet_v2_predictions.tar.gz	NCST predictions for 116 trained runs and five perturbation conditions, including the ViT and EduYOLO comparisons
rafnet_v2_scb3u.tar.gz	SCB3-U features, 15 ConvNeXt/Gated/RAFNet checkpoints and 16 comparison prediction runs
SHA256SUMS.txt	SHA-256 hashes of the four archives

See paper reproduction for download, training, inference and summary commands.

Raw images, original filesystem paths, classroom names, recording dates, and third-party encoder/SMPL weights are excluded. Opaque IDs preserve the image/group partitions. The feature arrays, labels and sample order are unchanged. The image-trained ViT/EduYOLO comparisons are supplied as predictions; their image-training checkpoints and private image data are not in the cached-feature archive. Restored-image controls have predictions but their expanded training features are not included.