arwics/cifar10

CIFAR-10, flattened to CSV

The first 5,000 rows of the CIFAR-10 train split as raw pixels, one row per image — a pinnable CSV for the Bend diffusion experiments in adeerkhan/MLBend (examples/diffusion_digits's 3072-output sibling). No CSV of CIFAR-10 existed on the Hub, so this is derived deterministically from uoft-cs/cifar10.

Format

• 5,000 rows × 3,073 columns: 3,072 pixel values then the label.
• No header, comma-separated — the mstz/optdigits convention that mlbend-datasets' csv_of(repo, rev, "train.csv", False{}, 0n) reads directly.
• Pixels are integers 0..255, channel-major: the 1024 red values, then 1024 green, then 1024 blue (the PyTorch CHW order — a PyTorch comparison is a reshape). Divide by 255.0 for [0, 1].
• Label: 0..9 (airplane, automobile, bird, cat, deer, dog, frog, horse, ship, truck), last column.

Provenance (pinned and reproducible)

Source, pinned by revision:

• dataset: uoft-cs/cifar10
• revision: 0b2714987fa478483af9968de7c934580d0bb9a2
• file: plain_text/train-00000-of-00001.parquet
• parquet sha256: 8428b53a88a11ac374111006708df51469e315a22ac6d66470afd9c78d2ae883

Derivation (deterministic, lossless): decode each of the first 5,000 PNGs (32×32 RGB, lossless), transpose to CHW, write ints and label.

This file:

• bytes: 54,902,548
• sha256: 80f3ca81e56946a0eabc65f68b69558b4fdc741e1aafb0d9f552fcee768b3161

Verified: row 0 read back from the CSV equals the channel-major form of the source PNG element-for-element; all values in 0..255; all ten labels present.

Usage in MLBend

"dataset": {"name": "cifar10", "rows": 5000, "features": 3072,
            "source": "https://huggingface.co/datasets/arwics/cifar10",
            "files": ["https://huggingface.co/datasets/arwics/cifar10/resolve/<revision>/train.csv"],
            "expected_sha256": "80f3ca81e56946a0eabc65f68b69558b4fdc741e1aafb0d9f552fcee768b3161"},