A TUI for Prodigy, made with Textual.
You can watch the brief overview demo below here:
prodigy-tui-yt.mp4
This project is currently mainly meant as a demo. Partially because doing annotations via the terminal is relatively experimental, but also because I cannot guarantee full feature parity with Prodigy just yet. The only implemented interface right now is textcat.manual but I may implement more in the future.
That said, this tool does integrate with Prodigy. For example, when you run this:
> python -m prodigy-tui textcat.manual --help
You're going to get a very familiar help text.
usage: textcat.manual [-h] [--label LABEL] dataset source
Interface for binary text classification from the terminal!
positional arguments:
dataset dataset to write annotations into (str)
source path to text source (Path)
optional arguments:
-h, --help show this help message and exit
--label LABEL, -l LABEL
category label to apply, only binary is supported (str)
To start annotating binary text examples, you can run something like:
prodigy-tui textcat.manual reddit-data examples.jsonl --label positive
And this will start saving annotations into the reddit-data dataset.
You will need to install this tool in an environment that already has Prodi.gy installed. Given such an environment, you can use pip.
python -m pip install prodigy-tui
To get the latest from Github, you can also use:
python -m pip install "prodigy-tui @ git+https://github.com/koaning/prodigy-tui.git"
- Saves the data to Prodigy like normal, keeping config files in mind.
- Works over SSH, no need to share a browser.
- Keyboard shortcuts are the same.
- You could even click the buttons if you really wanted to though.
- Support highlights in text
- Support multiple choices (inclusive/exclusive)
- Support images
- Add hierarchy between components/cards
As mentioned before, this is a bit of an experiment. But it seems interesting to explore. Maybe data annotation is something that can happen via the terminal and maybe there's something that's very developer friendly about that.
Props to the folks over at textualize. Much of the heavy lifting in UI-land is handled by their stack.