Gradygu3u's picture
upload mindcube_framework/metadata: metadata/README_TRANSFER.md
dfb6375 verified
|
Raw
History Blame Contribute Delete
2.04 kB

Spatial Data Transfer Inventory

Generated: 2026-06-12 20:25:33 Server root: /home/tione/notebook/tianjungu

This is a handoff package for transferring the spatial data/work directories. It does not copy TB-scale videos into this folder; it records exact source paths and ready-to-use rsync lists.

Recommended transfer order

  1. Transfer core manifests and release archives with rsync_commands_example.sh.
  2. Restore layout using restore_training_layout.sh from the release package.
  3. Use cambrianp_train_manifest_media_unique.tsv and escher_video_list_resolved.tsv to verify media availability.
  4. Use loose-file lists only for patch transfer or spot repairs; full loose transfer is slower than archives.

Important files

  • topic_video_summary.tsv: counts, sizes, and samples by topic.
  • project_and_data_paths.tsv: project/data roots to understand the layout.
  • escher_video_list_resolved.tsv: every Escher video_list entry mapped to an actual mp4 path.
  • cambrianp_train_manifest_media_unique.tsv: unique image/video references from Cambrian-P manifest, with logical path and resolved actual path.
  • cambrianp_image_folder_symlink_map.tsv: symlink map for Cambrian-P media layout.
  • release_raw_archives_inventory.tsv: all raw archive files in the release package.
  • files_to_transfer__*.absolute.txt: human-readable absolute source paths.
  • files_to_transfer__*.rsync_from_root.txt: rsync-ready path lists, used with source /.
  • rsync_commands_example.sh: conservative transfer template.

Critical note

Cambrian-P train media is a symlinked logical layout. Copying only /home/tione/notebook/tianjungu/datasets/Cambrian-P-train-data is insufficient unless the symlink targets are also transferred or reconstructed:

  • /home/tione/notebook/tianjungu/hf_cache/VSI-590K
  • /home/tione/notebook/tianjungu/datasets/Cambrian-S-3M

Geothinker/SSR data directories do not directly store videos in this inventory; they point to annotations/models/structured data and depend on the media roots above.