zevatov/nra-cifar10

🧬 CIFAR-10 in NRA Format — Zero-Download Cloud Training

This dataset contains CIFAR-10 (60,000 images) packaged in the NRA (Neural Ready Archive) format — a next-generation binary format built in Rust for the AI era.

🚀 Why This Matters

You DO NOT need to download this dataset. NRA streams data directly into your PyTorch DataLoader via HTTP Range requests. Only the exact 4MB blocks your model needs are fetched on-the-fly.

Metric	Traditional (tar.gz)	NRA (this dataset)
Time to first batch	~30 min (download + unpack)	150 ms
Local disk space	170 MB	0 bytes
Random file access	Impossible	O(1) instant

---

⚡ Quick Start: Train in 30 Seconds

Google Colab / Jupyter / Local

pip install nra==1.0.2 torch

import nra
import torch
from torch.utils.data import Dataset, DataLoader

class NraStreamDataset(Dataset):
    def __init__(self, url):
        self.url = url
        # The manifest downloads in ~150ms. The archive stays on Hugging Face!
        self.file_ids = nra.CloudArchive(url).file_ids()
        self._archive = None
        
    def __len__(self):
        return len(self.file_ids)
        
    def __getitem__(self, idx):
        if self._archive is None:
            self._archive = nra.CloudArchive(self.url)
            
        file_id = self.file_ids[idx]
        
        # NRA fetches only the exact chunk via HTTP Range.
        # The GIL is released; Rust streams data at max speed.
        raw_bytes = self._archive.read_file(file_id)
        
        # For real training: decode the image
        # img = Image.open(io.BytesIO(raw_bytes))
        # tensor = transforms.ToTensor()(img)
        return torch.tensor([len(raw_bytes)], dtype=torch.float32)

# Point directly to the .nra file in this repository
dataset = NraStreamDataset(
    "https://huggingface.co/datasets/zevatov/nra-cifar10/resolve/main/cifar10.nra"
)
loader = DataLoader(dataset, batch_size=256, num_workers=4)

print(f"✅ Loaded {len(dataset)} items. Training starts NOW — zero bytes on your SSD!")

for batch in loader:
    # Your model trains immediately. No waiting, no downloading.
    pass

---

🛠️ CLI: Inspect, Stream, or Mount

If you prefer working from the terminal:

# Install the Rust CLI
cargo install nra-cli

# Stream a single file without downloading the archive
nra-cli stream-beta \
  --url https://huggingface.co/datasets/zevatov/nra-cifar10/resolve/main/cifar10.nra \
  --file-id image_001.png \
  --out ./image_001.png

# Mount the remote archive as a local folder (Mac/Linux FUSE)
nra-cli mount \
  --input https://huggingface.co/datasets/zevatov/nra-cifar10/resolve/main/cifar10.nra \
  --mountpoint ./virtual_dataset

# Your files appear as a regular folder — but they're streaming from Hugging Face!
ls ./virtual_dataset/

---

🏗️ How It Works

PyTorch DataLoader → NRA Core (Rust) → HTTP Range GET → Hugging Face CDN
                                                              ↓
                                              Only the 4MB block you need
                                                              ↓
                                              Zstd decompress in RAM
                                                              ↓
                                              Raw bytes → GPU tensor

NRA uses:

• B+ Tree Manifest for O(1) file lookups (no scanning)
• 4MB Solid Blocks with Zstd compression
• HTTP Range Requests to fetch only the exact bytes needed
• Content-Defined Chunking (CDC) for automatic deduplication

---

🔄 Convert Your Own Datasets

Have a tar.gz or zip dataset? Convert it to NRA in seconds:

# Unpack and repack as NRA
nra-cli pack-beta --input ./your_dataset/ --output your_dataset.nra --dictionary --zstd-level 15

# Upload to your own HF dataset
# Then use the same streaming code above with your URL!

---

📊 Dataset Details

Field	Value
Source	CIFAR-10 (Krizhevsky, 2009)
Format	.nra (Neural Ready Archive v4.5)
Images	60,000 (32×32 RGB)
Classes	10
Compression	Zstd (level 15) + CDC deduplication
NRA SDK	pip install nra==1.0.2

---

📚 Learn More

• 🏠 GitHub Repository — Full source code, benchmarks, whitepapers
• 📦 PyPI Package — pip install nra==1.0.2
• 📄 Technical Whitepaper — Architecture deep-dive with 8 benchmark charts

License

This dataset and the NRA format are released under the MIT License.