VA-data — lingbot-va research corpus + measurement records

A machine-handover bundle: everything outside the git repo that the AgiBot-rehearsal line needs, tarred (no compression) so the Hub sees ~280 large files instead of millions of small ones.

Restore with RESTORE.sh (downloads, reassembles split units, untars to the ORIGINAL absolute paths, recreates the eval symlinks and the export→base-model symlinks):

export HF_ENDPOINT=https://hf-mirror.com
hf download Mo-ZheHan/VA-data --repo-type dataset --include "RESTORE.sh" --local-dir .
bash RESTORE.sh 01 02 03        # records + eval instrument + r7b continuity first
bash RESTORE.sh all             # everything

Groups, in restore-priority order

group	what	size	replaceable?
01-records	eval cell cache (1471 cells = every offline TF/vdrift number), SR closed-loop results, setup_logs diagnostic toolchain, the full git repo incl. history, other eval-cell groups	~14G	no — re-deriving needs 200G+ of checkpoints and GPU-days
02-eval-instrument	the two RoboTwin eval tasks WITH videos (eval decodes mp4; training never does)	3.5G	only by re-collecting
03-r7b	r7b resume checkpoint (step 3000) + its step2000 export	69G	yes, by retraining ~11h
04-clean500	clean base corpus, 50 task tars (latents + text_emb)	113G	slow re-extraction
05-geniesim	filtered GenieSim DR rehearsal, 18 task tars, videos included (the only re-extraction source)	129G	slow
06-baseline-exports	r2 + r7 exports at 2k–10k — the baselines any new eval must pair against	108G	re-exportable from full ckpts (not uploaded)
07-agibot	AgiBot rehearsal corpus, 188 task tars	242G	slow re-fetch
08-base-model	lingbot-va-base	23G	yes — hf download robbyant/lingbot-va-base

MANIFEST.tsv lists every uploaded path with its byte size; LFS sha256 is checked by hf download itself.

Deliberately NOT here

• Mix trees (datasets/mix/*): build_mix_tree trims metadata only, so a tree still hardlinks every episode of every task it includes — copying one moves 220G+ for nothing. Rebuild them (commands in RESTORE.sh's closing notes).
• geniesim/lerobot_v21: same inodes as the videos inside 05-geniesim.
• Full training checkpoints (21G each): wandb holds the curves, the cells hold the measurements.
• .venv: must be rebuilt on local disk, never on a network share.
• Credentials (~/.netrc, HF token, ssh keys/config): this repo is public — carry those by hand, out of band.