didfd's picture
fix: harden verified restore workflow
0821f99 verified
|
Raw
History Blame Contribute Delete
4.16 kB
---
license: apache-2.0
pretty_name: RoboDojo Assets (Packed)
tags:
- robotics
- simulation
- robodojo
- archive
---
# RoboDojo Assets — verified transport shards
This public dataset is a byte-preserving transport repack of `Assets/` from
[`RoboDojo-Benchmark/RoboDojo`](https://huggingface.co/datasets/RoboDojo-Benchmark/RoboDojo), pinned at source commit
`aada55a7a58b520a2004b9db02e884144b6e83e7`. It reduces 15,335 source files to 11 independently downloadable and
restorable `tar.zst` shards. No payload file was changed.
这是 RoboDojo `Assets/` 的公开传输分片:固定源 commit、减少小文件数量、每片可独立下载/重试,
完整复原后会逐文件做 SHA-256 校验。
## Contents
- Source files: **15,335**
- Source logical bytes: **41,268,637,776**
- Source revision: **`aada55a7a58b520a2004b9db02e884144b6e83e7`**
- Archive format: GNU tar + Zstandard level 3
- Every archive is below **4,000,000,000 bytes**
| shard | files | input bytes | archive bytes | SHA-256 prefix |
|---|---:|---:|---:|---|
| `data/assets-000.tar.zst` | 2,023 | 3,758,095,437 | 3,538,570,683 | `a0846895771da2ac…` |
| `data/assets-001.tar.zst` | 143 | 3,758,074,568 | 3,749,841,773 | `711b3a5bcee029cd…` |
| `data/assets-002.tar.zst` | 263 | 3,758,095,164 | 3,175,194,388 | `b8360b8f0bf21af0…` |
| `data/assets-003.tar.zst` | 491 | 3,758,080,869 | 3,488,084,115 | `9b184cbe5ccc249e…` |
| `data/assets-004.tar.zst` | 143 | 3,758,095,598 | 3,672,499,179 | `442e12ad6c7820aa…` |
| `data/assets-005.tar.zst` | 472 | 3,758,095,587 | 3,385,701,591 | `1e5b0fdf922b09e7…` |
| `data/assets-006.tar.zst` | 467 | 3,758,084,191 | 3,491,925,802 | `1469148d3284b53b…` |
| `data/assets-007.tar.zst` | 8,488 | 3,758,059,501 | 3,311,607,868 | `29d4b73bf90de4d3…` |
| `data/assets-008.tar.zst` | 517 | 3,758,096,082 | 3,433,188,150 | `045b896487b1484c…` |
| `data/assets-009.tar.zst` | 592 | 3,758,062,850 | 3,570,379,301 | `0a7aa894670f58e9…` |
| `data/assets-010.tar.zst` | 1,736 | 3,687,797,929 | 3,459,212,794 | `7e4d65469a712e59…` |
`SHA256SUMS` verifies the 11 transport archives. `FILES.sha256` verifies every restored
source file. `MANIFEST.json` records exact counts, sizes, paths, and full hashes.
## Download with low external concurrency
Install the current Hugging Face CLI, then choose one or two workers. Pin `--revision`
to a full commit SHA in production.
```bash
hf download didfd/robodojo-assets-packed \
--repo-type dataset \
--revision <FULL_PACKED_REPO_COMMIT_SHA> \
--local-dir /path/to/robodojo-assets-packed \
--max-workers 2
```
If the company importer accepts file URLs only, import the 11 files under `data/` plus
`MANIFEST.json`, `SHA256SUMS`, `FILES.sha256`, `lists/`, and `restore.py`. With concurrency
1, only one ~3.7 GB object is active at a time; a failed transfer only retries that shard.
## Restore and verify
Python 3.9+, `zstd`, and Linux with `/proc/self/fd` plus `O_NOFOLLOW` support are required.
The restore tool validates all metadata, archive hashes, and structured tar headers before
creating the output. It rejects links, sparse files, devices, FIFOs, non-canonical paths,
duplicates, and cross-shard path conflicts. Each shard is extracted into an isolated private
staging directory; only a fully hashed tree is atomically published as `Assets/`.
```bash
python3 restore.py \
--repo-dir /path/to/robodojo-assets-packed \
--output-dir /path/to/restored \
--workers 4 \
--hash-workers 16
```
Success ends with `RESTORE VERIFIED`; the restored root is
`/path/to/restored/Assets/`. If `Assets/` already exists, `--force` builds and verifies a
fresh tree first, atomically replaces the old tree, and preserves the old tree as
`Assets.backup.<random-id>`; it does not perform an in-place overwrite. Atomic replacement
requires filesystem support for Linux `RENAME_EXCHANGE`; if unavailable, `--force` fails
without modifying the existing tree.
## Provenance and license
The source dataset declares Apache-2.0. This repository changes packaging only and keeps
the same declared license; users remain responsible for following the upstream dataset's
terms and notices.