| # Restore the original archive files | |
| The repository is being uploaded. Each original listed in `manifests/` has a complete set of binary pieces. Other originals may still be pending. | |
| Download the pieces, `manifests/`, and `restore_archives.py` to one directory. Once your access request has been approved and you have logged in to Hugging Face: | |
| ```bash | |
| hf download lanhe1124/archives --repo-type dataset --local-dir archives | |
| cd archives | |
| python3 restore_archives.py | |
| ``` | |
| To restore just one available original: | |
| ```bash | |
| python3 restore_archives.py --file gcd_kubric4d_test.tar.gz | |
| ``` | |
| The script verifies every piece's size and SHA-256, concatenates pieces in manifest order, and verifies the complete original SHA-256. It never overwrites an existing original or deletes pieces. Restoration requires additional free space equal to the original size. It does not decompress archives. | |
| To write restored files elsewhere, add `--output /path/to/output`. After a failed or interrupted attempt, inspect and remove only its `.restoring` output before restarting; existing completed originals are preserved. | |
| The training originals named `.aa` through `.ar` are themselves volumes of the original training archive. Restoring their `.partNNN` pieces reproduces those original volumes byte for byte. | |