You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
Access to this research backup requires manual approval by the owner. Original dataset licenses and restrictions continue to apply.
Log in or Sign Up to review the conditions and access this dataset content.
Research machine data backup — 2026-10-07
Backup in progress. Do not treat this repository as a completed machine backup.
This dataset preserves locally created annotations, datasets actually used in research, and selected experiment artifacts before the source machine is cleared. Access is gated with manual approval. This is a recovery archive, not a normalized benchmark release.
Data archives are grouped by source and priority. Each .tar.zst is independently
extractable and retains the original path below a restoration directory. The inventory
and upload receipts will identify coverage, exclusions, unavailable paths, and revisions.
Source licenses are retained; this backup does not grant additional redistribution rights.
Credentials and machine authentication files are excluded. Any credential redactions
in archived text will be documented in the manifests.
Code is backed up separately to the owner's private repository: https://github.com/shulin16/machine-code-backup-20261007
The requested second copy is Google Drive folder
1dOyGIeGGlVTzIANRROSGJzZsNuDOL5os; its upload is awaiting Google OAuth authorization.
Restore an archive
After the owner approves access, download the required files using hf download.
Extract into a new directory first, without overwriting an existing machine:
mkdir -p restored
tar --zstd -xf ARCHIVE.tar.zst -C restored
For example, a member named data/robomme/annotations/... is restored below
restored/data/robomme/annotations/.... Update absolute dataset paths or recreate
the original directory layout when resuming training. Review the inventory for symlink
targets, externally hosted upstreams, and paths not captured in this backup.
Completed annotation coverage
All 205,205 inventoried annotation metadata entries, 1,120,359 review images,
and four historical cache TAR paths have been preserved. The current 39 data archives
contain 204,962 metadata entries and occupy 48,542,762,369 compressed bytes;
243 historical project source files are retained in three verified private original
metadata archives. The combined metadata paths and original byte counts exactly
match the initial inventory. Three distinct cache TARs are stored; one exact duplicate
has a restore mapping. All 39 current data archives were decompressed and their member
inventories checked; their bytes match the Hub's native LFS checksums. The coverage
check found no missing, changed, or unreadable source entries. See
manifests/annotation_remote_verification.json and the annotation recovery bundle.
Other local datasets and experiment archives continue uploading.
The annotation recovery bundle includes annotation_private_source_routing.json,
which lists the private originals and their source members. Restore these originals
before applying the current annotation data archives. The complete working project
code and Git histories are also in the private GitHub backup.
Some experiment metadata archives originally included project source files. The full
original archives are now retained as private GitHub release assets named
source-evidence--*.tar.zst.part000; current data copies omit those source members.
Each source-evidence asset is one complete, independently extractable tar.zst stream.
Do not concatenate different source-evidence assets. For a combined restoration,
extract private originals in the capture order recorded in
resume-ordinary-source-private-copies.jsonl, then overlay the current data archives.
The migration receipts and code-evidence-private.jsonl record these source mappings.
Historical Hugging Face versions are still awaiting privacy cleanup; use the current
repository files for restoration.
source-repositories/ contains pinned original upstream copies, verified against their
existing Hub checksums. Local modifications and additions are archived separately.
An upstream copy alone is not claimed as proof of local-file equivalence.
Recovery manifests during the upload
manifests/data_recovery_latest.json points to the newest verified data recovery
metadata bundle. Download and extract that bundle to recover inventories, per-archive
member lists, exclusions, source mappings, upload receipts, and available multipart
reconstruction maps. Each snapshot records its capture time and explicitly remains
in progress until the overall backup is reconciled. Per-archive manifests describe
completed local archives; upload receipts identify which had reached the Hub at
the snapshot time. Annotation recovery metadata is published in its separate bundle.
Archives retain original paths under the chosen restoration root. If an individual source file was split, extract every listed part first, then run:
python tools/restore_large_files.py --restore-root restored \
--manifest manifests/data_large_file_parts.jsonl
The multipart manifest is present only after at least one split source file has completed. This command expects the parts for every listed source; for a partial restore, use a copy of the manifest containing only the required source rows.
The local LeRobot dataset was independently checked with the native Hub verifier:
all 1,606 upstream files (129,485,389,810 bytes) match pinned revision
1510653cccb4d9e5165fb3141c06d88053decc20. Restore them from
source-repositories/Yinpei--robomme_data_lerobot/, then apply
archives/robomme-lerobot-overlay/ for the local addition. The mapping is recorded
in manifests/verified_upstream_local_mappings.json.
Two more verified local subsets use owned source mirrors:
| Local data | Owned mirror path | Exact restore selection |
|---|---|---|
| Franka compressed demonstrations: 158 H5 files, 85,418,738,152 bytes | source-repositories/fm-dev--franka-demos-compressed/ |
franka-compressed-verified-h5.jsonl; apply franka-compressed-local-metadata archives afterward |
| RoboDojo: 731 USDZ assets, 12,600,034,910 bytes | source-repositories/RoboDojo-Benchmark--RoboDojo/ |
source_path entries in resume-audit-robodojo-container-mirror.jsonl; apply robodojo-referenced-simulator-assets archives for other assets and two local YAML overlays |
The recovery bundle includes these path lists, native verification receipts, and
verified_upstream_local_mappings.json, which supplies the original local roots.
Copy each selected mirror path to the same relative location beneath that restored
root. Upstream mirrors and local archives can overlap, so their byte totals are not
independent unique-data totals.
Latest Franka all-task training checkpoints
The four latest step-6250 checkpoint families (pick3/shuffle, each with
baseline/uniform32) have a combined restore selection of 135 source members,
including 48,739,015,669 regular-file bytes. Their complete training state reuses
25 archives across the prepared-data and latest-checkpoint groups. Download every
archive listed in manifests/franka-alltask-latest-restore-selection.json before
restoring these families; downloading only the newest checkpoint group is insufficient.
franka-alltask-latest-restore-files.jsonl maps each selected source member to its
archive. franka-alltask-latest-remote-verification.json records the native checks
for all 25 required archives. These files are in the recovery metadata bundle.
Crossbench canonical H5 paths
During this backup, RoboMemArena preparation moved completed H5 files from
prepared-full/work/shard-*/episodes/ into prepared-full/episodes/. The fixed
canonical inventory has 2,600 H5 files totaling 585,916,341,860 bytes. Captures at
the older work paths and the crossbench-final-h5-supplement archives together
provide the recovery inputs. The supplement alone does not contain the full set.
crossbench-final-h5-restore-files.jsonl maps a verified archive member to its
canonical restore_path; restore it below a new recovery root. A mapping appears
only after an exact byte comparison with the canonical file, or a stable direct
capture at the canonical path. crossbench-final-h5-restore-selection.json lists
required archives and pending paths. This reconciliation is still in progress;
its partial mapping is not proof that all 2,600 files have been preserved.
Native remote archive verification remains a separate check.
The crossbench-final-feature-metadata group preserves the finalized episode
inventory and global feature statistics. Live writer/status sidecars and final
training transport packages require their own later inventory and verification.
- Downloads last month
- 139