ScanData

Audit your ML dataset before you train on it.

Point ScanData at a dataset folder and it tells you what will break your training: leakage between splits, label problems, shortcuts the model can cheat with, and image quality issues. Every finding comes with a severity, evidence and a concrete fix.

Early preview. This release ships the interactive CLI and the full check catalog. Dataset scanning is coming in the next release.

Install

pip install scandata

Requires Python 3.10 or newer. Works on Windows, macOS and Linux.

Quick start

Run it with no arguments to open the interactive menu:

scandata

   ____                  ____        _
  / ___|  ___ __ _ _ __ |  _ \  __ _| |_ __ _
  \___ \ / __/ _` | '_ \| | | |/ _` | __/ _` |
   ___) | (_| (_| | | | | |_| | (_| | || (_| |
  |____/ \___\__,_|_| |_|____/ \__,_|\__\__,_|

  ScanData · audit your dataset before you train on it
  Built with ❤ by Subham Divakar

› What would you like to do?
  » 1. Scan a dataset
    2. Results (last scan)
    3. Browse checks
    4. Recent scans
    5. Settings
    6. Help & about
    0. Exit

Or skip the menu and run commands directly, which is handy in scripts and CI:

scandata scan ./data --type image-cls          # scan a dataset folder
scandata list-checks                           # every check, in a table
scandata list-checks --section leakage         # one section only
scandata explain leak.near_dup                 # what a check does and why it matters
scandata --version
scandata --help

What it checks

Checks for image classification datasets, grouped by what they protect you from:

Run scandata list-checks for the full list, or scandata explain <check-id> for details on any one.

Supported dataset layouts

class folders               split / class folders        manifest
data/                       data/                        data/
  cracked/   *.jpg            train/cracked/ ...           images/...
  uncracked/ *.jpg            val/...                      labels.csv
                              test/...                     (path,label[,split][,group])

ScanData works out the layout automatically.

Scan options

Exit codes: 0 OK, 1 the --fail-on threshold was hit, 2 error.

Contributing

Contributions are welcome, whether it's a bug report, a new check idea, docs or code. Feel free to reach out: open an issue or contact me through GitHub.

To contribute code:

1. Fork the repo and clone your fork:

   git clone https://github.com/<your-username>/scandata.git
   cd scandata
   

2. Create a virtual environment and install in editable mode with the dev tools:

   python -m venv .venv
   source .venv/bin/activate        # Windows: .venv\Scripts\activate
   pip install -e ".[dev]"
   

3. Create a branch for your change:

   git checkout -b my-change
   

4. Make your change, then run the tests and the linter:

   pytest
   ruff check src tests
   

5. Push your branch and open a pull request describing what you changed and why.

For larger changes, such as a new check or data type, please open an issue first so we can agree on the approach.

License

MIT

Built with ❤ by Subham Divakar.