# nwm-ogb-cube-off-v5
Rig-v5 (ghost-free) multiview NVS set rendered from the OFFICIAL OGBench `cube-single-play-v0` play dataset
(1000 train episodes x 1000 steps; eval from `cube-single-play-v0-val`). 9 views/scene (5 ctx + 4 gap-midpoint
targets, nv_arc=5), 224px, GT depth + seg. train = 5 states/episode, fixed permutation (5000 scenes);
eval = one state from each of 30 val episodes. ds_idx.npy = (episode, absolute row) into the official npz.
Supersedes nwm-ogb-cube-v5 (which replayed the swm galilai-group/ogb_cube_single set — a different,
201-step dataset). Tokenizer trained here: Y31-ogb.
Files: s{k}_v{j}.png, s{k}_depth.npy, s{k}_seg.npy, poses.npz, ds_idx.npy, meta.json, labels.json.

Layout: train.tar / eval.tar (untar to use; frozen s*_v*.png format inside).