narteybrown/signal-bench-ic-v1

Signal-Bench IC Test Set v1

What this is

The CIFAR-10 test set, packaged for use by signal-bench — an open-source benchmarking framework for edge AI hardware. This dataset is the evaluation set used in Phase 5 image-classification benchmarks across the MCU-to-cloud hardware curve.

This is not a new dataset. It is a packaging of CIFAR-10's standard test split with no modification, published in HuggingFace's datasets format for one-line loading.

Source attribution

CIFAR-10 was collected by Alex Krizhevsky, Vinod Nair, and Geoffrey Hinton at the University of Toronto. The original dataset is available at https://www.cs.toronto.edu/~kriz/cifar.html.

Citation:

Krizhevsky, Alex. "Learning multiple layers of features from tiny images." (2009).

Processing

Passthrough. CIFAR-10 test images are stored as-is at 32×32×3 uint8 (RGB). No normalization, no cropping, no augmentation. The reference MLPerf Tiny image classification model (pretrainedResnet_quant.tflite) applies its own normalization internally during inference.

Subset structure

• Samples: 10,000 (full CIFAR-10 test set)
• Classes: 10 (airplane, automobile, bird, cat, deer, dog, frog, horse, ship, truck)
• Class balance: 1,000 per class
• Feature shape: 32×32×3 uint8 RGB

Intended use

Benchmark evaluation for edge AI image classification. Designed to pair with MLPerf Tiny's reference ResNet-8 model.

from datasets import load_dataset
ds = load_dataset("narteybrown/signal-bench-ic-v1")
print(ds)  # DatasetDict with single 'train' split (HF convention; semantically a test set)

The split is named train per HuggingFace's default convention for single-split datasets, but it is semantically a test set. Do not use it for model training.

MCU subset

A deterministic 100-sample stratified subset (10 samples per class, seed=42) is committed to the signal-bench repo at data/mcu_subsets/ic/subset_v1.npz for flashing into MCU firmware as C arrays. To regenerate: python scripts/datasets/prepare_ic.py.

Citation

If you use this dataset in published work, cite both the original CIFAR-10 paper and signal-bench:

@misc{signalbench-ic-v1,
  author = {Brown, Daniel},
  title = {Signal-Bench IC Test Set v1},
  year = {2026},
  publisher = {Agoo AI},
  howpublished = {\url{https://huggingface.co/datasets/narteybrown/signal-bench-ic-v1}},
}

Limitations

CIFAR-10 is a small, well-studied dataset with known limitations: low resolution, limited class diversity, no environmental variation. Benchmark results using this set should be understood as a measurement of inference performance on a standard reference workload, not as evaluation against real-world deployment conditions.

The MCU subset (100 samples) is small enough that per-class accuracy estimates have wide confidence intervals; it is intended for latency and energy measurement, not statistical accuracy reporting.

License

CIFAR-10 is distributed under the terms documented at https://www.cs.toronto.edu/~kriz/cifar.html. This packaging adds no restrictions; users should consult the original license for any redistribution.