kaldi110/downloads-archive

Downloads Archive

Directories archived as ZIP files, each split into 500 MiB parts so they could be produced and uploaded on a machine with less free disk than the data.

Folder	Contents	Raw size
vscode/	VS Code + WinSCP installers, Remote-SSH .vsix extensions	~237 MiB
poppler/	Poppler 26.09.0 Windows release zip + upstream source tarball	~44 MiB
GPT/	Local GPT serving bundle, incl. gpt-oss-20b-MXFP4.gguf (~12 GiB)	~13 GiB
DiffSinger/	DiffSinger / OpenUtau / MFA trees: checkpoints, venvs, voicebanks	~37 GiB

Layout

Each folder holds the parts of one ZIP:

vscode/vscode.zip.001
poppler/poppler.zip.001
GPT/GPT.zip.001, GPT.zip.002, ...
DiffSinger/DiffSinger.zip.001, DiffSinger.zip.002, ...

A folder that fits in one part still uses the .001 suffix, so the same concatenate-then-unzip steps work everywhere.

The parts are a raw byte split of a single ZIP stream — not a multi-volume ZIP. Concatenate them in order and you get the original archive back. Every part except the last is exactly 524288000 bytes.

manifest.json lists every part with its size and SHA-256.

Reassembly

Download a folder, concatenate its parts, then unzip.

Linux / macOS

huggingface-cli download kaldi110/downloads-archive --repo-type dataset \
  --include "DiffSinger/*" --local-dir .
cat DiffSinger/DiffSinger.zip.* > DiffSinger.zip
unzip DiffSinger.zip

Windows (PowerShell)

huggingface-cli download kaldi110/downloads-archive --repo-type dataset `
  --include "DiffSinger/*" --local-dir .
cmd /c "copy /b DiffSinger\DiffSinger.zip.* DiffSinger.zip"
Expand-Archive DiffSinger.zip -DestinationPath .

Python, without a full local copy of the joined file

import shutil, glob
with open("DiffSinger.zip", "wb") as out:
    for part in sorted(glob.glob("DiffSinger/DiffSinger.zip.*")):
        with open(part, "rb") as f:
            shutil.copyfileobj(f, out)

Verifying

python - <<'EOF'
import json, hashlib, pathlib
m = json.load(open("manifest.json"))
for path, meta in m["parts"].items():
    p = pathlib.Path(path)
    if not p.exists():
        continue
    h = hashlib.sha256()
    with open(p, "rb") as f:
        for b in iter(lambda: f.read(1 << 20), b""):
            h.update(b)
    print("OK " if h.hexdigest() == meta["sha256"] else "BAD", path)
EOF

Notes

• Archives use ZIP64 (GPT.zip contains a single ~12 GiB member), so extract with a ZIP64-capable tool. Windows Explorer's built-in extractor and older unzip builds can fail; 7-Zip or Python's zipfile work.
• Entries are stored uncompressed when the source was already a compressed format (.gguf, .ckpt, .zip, .exe, ...) and deflated otherwise.
• These trees contain third-party redistributables (NVIDIA CUDA installer, Microsoft VS Code installer, PyTorch wheels, UTAU/DiffSinger voicebank packs). They are covered by their own licenses, which are not granted by this repository. Check the upstream terms before redistributing any of it.