Blame for README.md

---
license: mit
task_categories:
  - image-classification
language:
  - en
tags:
  - mnist
  - ufo
  - uap
  - computer-vision
  - image-classification
  - synthetic-data
size_categories:
  - 10K<n<100K
pretty_name: UFO-MNIST
---

# UFO-MNIST

[![GitHub stars](https://img.shields.io/github/stars/tentime/ufo-mnist.svg?style=flat&label=Star)](https://github.com/tentime/ufo-mnist/)
[![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](LICENSE)
[![Dataset](https://img.shields.io/badge/dataset-28x28%20grayscale-blue.svg)](data/ufo)

<details><summary>Table of Contents</summary><p>

* [Why UFO-MNIST](#why-ufo-mnist)
* [Get the Data](#get-the-data)
* [Labels](#labels)
* [Usage](#usage)
* [Benchmark](#benchmark)
* [Visualization](#visualization)
* [Contributing](#contributing)
* [Citing UFO-MNIST](#citing-ufo-mnist)
* [License](#license)
</p></details><p></p>

`UFO-MNIST` is a dataset of UFO-like spotting patterns and common aerial
lookalikes, consisting of a training set of 8,000 examples and a test set of
2,000 examples. Each example is a 28x28 grayscale image associated with one of
10 labels.

UFO-MNIST is designed as a compact, MNIST-style benchmark for machine-learning
examples and image classifiers. It shares the original MNIST image size and
train/test structure, and it is provided both as a compressed NumPy archive and
as IDX gzip files compatible with common MNIST loaders.

Here's an example of how the data looks:

![](doc/img/ufo-mnist-sprite.png)

## Why UFO-MNIST

MNIST-style datasets are useful because they are small, fast, visual, and easy
to load. UFO-MNIST keeps those properties while moving away from handwritten
digits into low-resolution spotting categories: disks, orbs, triangles,
formations, glows, aircraft, balloons, birds, and celestial or sensor artifacts.

The dataset is assembled from public UFO/UAP sighting references, official
release material, and generated augmentations that make the classes balanced and
easy to use in MNIST-style experiments.

## Get the Data

You can use the NumPy archive directly:

| Name | Content | Examples | Size | Link | SHA-256 |
| --- | --- | ---: | ---: | --- | --- |
| `ufo_mnist_28x28.npz` | train/test images and labels | 10,000 | 5.7 MB | [Download](data/ufo_mnist_v1/ufo_mnist_28x28.npz) | `65a87cf1121c38c247862637faea5cb3d0381dcb18dc4dac42d9545f0ef6c5e6` |

The dataset is also stored in the same IDX gzip format used by the original
MNIST dataset:

| Name | Content | Examples | Link |
| --- | --- | ---: | --- |
| `train-images-idx3-ubyte.gz` | training set images | 8,000 | [Download](data/ufo/train-images-idx3-ubyte.gz) |
| `train-labels-idx1-ubyte.gz` | training set labels | 8,000 | [Download](data/ufo/train-labels-idx1-ubyte.gz) |
| `t10k-images-idx3-ubyte.gz` | test set images | 2,000 | [Download](data/ufo/t10k-images-idx3-ubyte.gz) |
| `t10k-labels-idx1-ubyte.gz` | test set labels | 2,000 | [Download](data/ufo/t10k-labels-idx1-ubyte.gz) |

Metadata:

| Name | Content |
| --- | --- |
| [`labels.json`](data/ufo_mnist_v1/labels.json) | label mapping |
| [`manifest.csv`](data/ufo_mnist_v1/manifest.csv) | source manifest |
| [`samples.csv`](data/ufo_mnist_v1/samples.csv) | per-sample split, label, source family, and seed |
| [`dataset_card.md`](data/ufo_mnist_v1/dataset_card.md) | dataset card |
| [`checksums.json`](data/ufo_mnist_v1/checksums.json) | release checksums |

## Labels

Each training and test example is assigned to one of the following labels:

| Label | Description |
| ---: | --- |
| 0 | disk |
| 1 | orb |
| 2 | triangle |
| 3 | cigar_rod |
| 4 | light_formation |
| 5 | irregular_glow |
| 6 | aircraft |
| 7 | balloon |
| 8 | bird |
| 9 | celestial_or_artifact |

## Usage

### Loading the NumPy archive

```python
import numpy as np

data = np.load("data/ufo_mnist_v1/ufo_mnist_28x28.npz")
X_train = data["train_images"]
y_train = data["train_labels"]
X_test = data["test_images"]
y_test = data["test_labels"]
```

### Loading the IDX files with Python

Use `utils/mnist_reader.py` in this repository:

```python
from utils import mnist_reader

X_train, y_train = mnist_reader.load_mnist("data/ufo", kind="train")
X_test, y_test = mnist_reader.load_mnist("data/ufo", kind="t10k")
```

### Build from source

```bash
python3 -m pip install -e .
ufo-mnist build --output data/ufo_mnist_v1 --seed 1337
ufo-mnist inspect --dataset data/ufo_mnist_v1
python3 scripts/export_idx.py
```

## Benchmark

The table below lists local benchmarks on the provided train/test split.

| Classifier | Preprocessing | Test accuracy | Macro F1 | Code |
| --- | --- | ---: | ---: | --- |
| Nearest centroid | raw pixels | 0.420 | - | [`src/ufo_mnist/inspect.py`](src/ufo_mnist/inspect.py) |
| Logistic regression | standardization | 0.459 | 0.456 | [`scripts/train_baseline.py`](scripts/train_baseline.py) |
| Small CNN | rescale to `[0, 1]` | 0.996 | 0.996 | [`scripts/train_cnn.py`](scripts/train_cnn.py) |

The CNN benchmark uses three convolutional blocks with batch normalization,
dropout, adaptive pooling, and AdamW. Full metrics are available in
[`cnn_metrics.json`](data/ufo_mnist_v1/cnn_metrics.json).

## Visualization

Training samples by class:

![](data/ufo_mnist_v1/contact_sheet_train.png)

Test samples by class:

![](data/ufo_mnist_v1/contact_sheet_test.png)

## Contributing

Issues and pull requests are welcome. Useful contributions include better
loaders, benchmark submissions, visualization notebooks, and reproducible model
scripts. If you submit a benchmark, include the exact train/test split, code,
seed, preprocessing, and test accuracy.

## Citing UFO-MNIST

If you use UFO-MNIST in a project or publication, cite this repository:

```bibtex
@misc{ufo_mnist_2026,
  title        = {UFO-MNIST: A 28x28 Grayscale Dataset of UFO-like Spotting Patterns},
  author       = {tentime},
  year         = {2026},
  howpublished = {\url{https://github.com/tentime/ufo-mnist}},
}
```

## License

MIT. See [LICENSE](LICENSE).