---
pretty_name: CNN Muffins
task_categories:
  - image-classification
tags:
  - dog
  - muffin
  - hard-negative
  - cnn
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files:
      - split: train
        path: train.jsonl
      - split: validation
        path: val.jsonl
  - config_name: hard16
    data_files:
      - split: test
        path: bench_hard16.jsonl
---

# CNN Muffins

A compact dog-versus-muffin image-classification dataset built around the
well-known visual confusion between Chihuahua faces and blueberry muffins.

![Canonical hard-16 dog-versus-muffin benchmark](hard16_4x4.jpg)

## Dataset structure

| Split | Dogs | Muffins | Total |
|---|---:|---:|---:|
| Train | 319 | 161 | 480 |
| Validation | 36 | 18 | 54 |
| Hard-16 benchmark | 8 | 8 | 16 |

The hard-16 benchmark is isolated from train and validation. The JSONL files
use repository-relative image paths:

The benchmark labels follow the original 4x4 checkerboard layout and were
corrected on September 2, 2026.

```json
{"image":"train/dog/example.jpg","label":"dog","dedupe_sha256":"...","file_sha256":"..."}
```

## Candidate filtering

Candidates were resized to 64×64 for comparison against the eight canonical
tiles of the opposite class. A candidate was retained when at least one
opposite-class reference satisfied all of the following:

- MSE < 0.17
- SSIM < 0.23
- HSV histogram intersection ≥ 0.72
- RGB mean/standard-deviation L2 distance ≤ 0.18

Retained candidates were deduplicated by SHA-256 before a deterministic,
stratified 90/10 train-validation split.

## Label-quality warning

The labels are **provisional and intentionally not cleaned**. A manual
sample-by-sample visual audit of all 534 train/validation images found:

- 427 correct labels
- 101 confirmed label errors (18.9%)
- 6 ambiguous images
- Dog-label confirmed error rate: 20.6%
- Muffin-label confirmed error rate: 15.6%

Some generic cat-versus-dog sources were assigned labels from path heuristics,
which introduced visible cat and out-of-class contamination. This release is
therefore suitable for robustness experiments, noisy-label research, and
pipeline debugging, but should not be treated as a clean supervised benchmark.
The canonical hard-16 benchmark labels are separate and were not included in
the 18.9% calculation.

## Intended use

- Training small CNN baselines without pretrained weights
- Measuring robustness to noisy labels and hard visual negatives
- Evaluating against the isolated canonical hard-16 benchmark

## Provenance and licensing

Images were aggregated from multiple public dataset sources. Upstream licenses
and usage terms may differ, and this repository does not grant additional
rights. Users are responsible for checking source-specific requirements before
redistribution or commercial use.