VLM-3R + VSI-590K + SPAR-7M-RGBD transfer bundle

Data only — no code, no model weights, no predictions. This is a byte-exact transport copy of three upstream datasets as they were staged on the source cluster, packaged so they can be moved to another machine and restored.

The three belong together: the VLM-3R training annotations reference SPAR frames, not VLM-3R's own media. See "How the pieces fit" below.

Contents

vlm3r/

File	Bytes	Contents
vlm3r_data.tar	213,053,440	The upstream Journey9ni/VLM-3R-DATA annotations: vsibench_train/ (3 JSON) and vstibench_train/ (11 JSON)
vlm3r_train_annotations.tar	2,355,937,280	vlm3r_vsi_205k{,_16frames,_32frames}.json and vlm3r_vst_132k{,_16frames,_32frames}.json — the training-format conversions actually consumed by training

vsi590k/

File	Bytes	Notes
vsi590k_annotations.tar	3,720,601,600	vsi_590k.jsonl, upstream README.md, and vsi_590k{,_16frame,_32frame}.json
adt.tar.gz	23,895,961,889
arkitscenes.tar.gz.part00 … part09	90,154,741,781	split — see below
hypersim.tar.gz	6,775,393,985
procthor.tar.gz	6,523,652,708
robotics.tar.gz	1,104,873,829
s3dis.tar.gz	2,865,916,732
scannet.tar.gz	27,632,592,849
scannetppv2.tar.gz	21,438,045,428
ytb_roomtour.tar.gz.part00 … part05	55,261,730,714	split — see below

spar7m_rgbd/

File	Bytes	Notes
spar-rgbd-00.tar.gz … spar-rgbd-17.tar.gz	184,377,341,136	one tar stream split across 18 parts — see below
spar-rgbd-sha256.txt	1,548	upstream's own checksums for the 18 parts
README.md	4,513	upstream dataset card

Total payload: 426,319,849,432 bytes.

All .tar.gz archives are the upstream ones, uploaded unmodified — they were not repacked. Only the loose annotation files were collected into tars.

Restore

SPAR-7M-RGBD

The 18 parts are a single tar stream split across files, not 18 independent archives. Extracting one part on its own will not work. Concatenate first, as upstream documents:

cd spar7m_rgbd
sha256sum -c spar-rgbd-sha256.txt
cat spar-rgbd-*.tar.gz | tar -xvzf -    # produces spar/

This yields spar/{scannet,scannetpp,structured3d,rxr}/{images,qa_jsonl}/.

VSI-590K

arkitscenes.tar.gz and ytb_roomtour.tar.gz exceed HuggingFace's 50 GB per-file limit, so each is transported as byte-exact 10 GB parts. Reassemble and verify those two first:

cat arkitscenes.tar.gz.part{00..09} > arkitscenes.tar.gz
sha256sum -c arkitscenes.tar.gz.original.sha256

cat ytb_roomtour.tar.gz.part{00..05} > ytb_roomtour.tar.gz
sha256sum -c ytb_roomtour.tar.gz.original.sha256

The remaining VSI-590K archives are independently extractable. Each unpacks into a per-source directory (adt/, arkitscenes/, …, ytb_roomtour/).

Everything else

MANIFEST.sha256 covers every uploaded object, including each split part individually, so it can be checked before any reassembly:

sha256sum -c MANIFEST.sha256
for archive in vlm3r/*.tar vsi590k/*.tar; do tar -xf "$archive"; done
for archive in vsi590k/*.tar.gz; do tar -xzf "$archive"; done

How the pieces fit

Restored into a project root, the expected layout is:

data/
  media/
    spar -> ../SPAR-7M-RGBD/spar     # symlink
  SPAR-7M-RGBD/spar/...              # from spar7m_rgbd/
  VSI-590K/{adt,arkitscenes,...}/    # from vsi590k/*.tar.gz
  VLM-3R-DATA/...                    # from vlm3r/vlm3r_data.tar
  train/*.json                       # from the two annotation tars

Two things are easy to get wrong:

• The VLM-3R training annotations do not use VLM-3R's own media. Every media reference in vlm3r_vsi_205k_32frames.json (205,456 rows) and vlm3r_vst_132k_32frames.json (132,053 rows) — 10,800,288 frame references in total — points at spar/…, resolved as data/media/spar/…. They need 64,860 distinct SPAR frames (8.36 GB) drawn from 1,201 ScanNet and 849 ScanNet++ scenes. Without spar7m_rgbd/ restored, these two files are unusable.
• VSI-590K frames resolve against a sibling root, not data/media. The annotations reference VSI-590k-128frames/<source>/<scene>/<scene>_frame_NN.jpg, which is looked up under data/VSI-590K/. When that pre-extracted frame directory is absent, the loader falls back to the parent video (data/VSI-590K/<source>/<scene>.mp4) and samples frames on the fly, which is what the vsi590k/*.tar.gz archives provide.

Upstream attribution

• VSI-590K — nyu-visionx/VSI-590K, revision 346fbd4e41dec974bf24894d0541a49327ee6669, Apache-2.0 dataset card. From Cambrian-S (Yang, Yang, Huang, Brown et al., arXiv:2511.04670). The underlying media is curated from ADT, ARKitScenes, Hypersim, ProcTHOR, S3DIS, ScanNet, ScanNet++ v2, robotics sources, and YouTube room-tour video, each of which retains its own upstream terms.
• VLM-3R — Journey9ni/VLM-3R-DATA, revision aae3535e513d36be5fd0a1edbaab523c8a3c45e4, Apache-2.0 dataset card. The copy here predates the 2026-07-13 camera-position erratum: five vstibench_train/qa_camera_*.json files are the larger pre-erratum versions and erratum_dropped_items.json is absent, because this is the revision the source project's experiments actually used. Do not silently replace it with the current upstream revision.
• SPAR-7M-RGBD — jasonzhango/SPAR-7M-RGBD, from the Fudan ZVG SPAR project (arXiv:2503.22976). Built over ScanNet, ScanNet++, Structured3D, and RxR scenes, which retain their upstream terms.

This mirror does not broaden or replace any upstream license, agreement, or attribution requirement. Cite the original datasets, not this bundle.