| DROID LeRobot dataset — split tar archive |
| ========================================= |
|
|
| The `droid_lerobot` dataset bundled as ONE uncompressed tar, split into 20 GiB |
| parts. This avoids transferring ~231k tiny files individually (which overwhelms |
| the Hub download); instead it's 7 large files. |
|
|
| Contents |
| -------- |
| droid_lerobot.tar.part00 .. part06 7 parts, ~130 GB total (last part smaller) |
| SHA256SUMS sha256 of each part, for integrity check |
| README_RECONSTRUCT.txt this file |
|
|
| Reconstruct (any environment with huggingface_hub installed) |
| ------------------------------------------------------------ |
| # 1. download all parts |
| hf download tturing/transfer --repo-type dataset --local-dir ./droid_transfer |
| cd droid_transfer |
|
|
| # 2. (optional) verify nothing corrupted in transit |
| sha256sum -c SHA256SUMS |
|
|
| # 3. reconstruct the dataset |
| mkdir -p /target/droid_lerobot |
| cat droid_lerobot.tar.part* | tar -xf - -C /target/droid_lerobot |
|
|
| This recreates data/, videos/, and meta/ under /target/droid_lerobot. |
|
|
| Notes |
| ----- |
| - No compression: the payload is already-compressed mp4 video + parquet, so |
| compression would cost ~1 hour of CPU for ~1% size reduction. |
| - The HuggingFace download cache (.cache/) was intentionally excluded — it is |
| local download bookkeeping and is not needed to use the dataset. |
| - The source had 4 unfinished (.incomplete) video downloads, so the dataset is |
| nearly complete but may be missing a handful of videos. |
|
|
