PromptSource

Toolkit for collecting and applying templates of prompting instances.

WIP

Setup

1. Download the repo
2. Navigate to root directory of the repo
3. Install requirements with pip install -r requirements.txt

Running

From the root directory of the repo, you can launch the editor with

streamlit run promptsource/promptsource.py

There are 3 modes in the app:

• Helicopter view: aggregate high level metrics on the current state of the sourcing
• Prompted dataset viewer: check the templates you wrote or already written on entire dataset
• Sourcing: write new prompts

Writing Templates

A prompt template is expressed in Jinja.

It is rendered using an example from the corresponding Hugging Face datasets library (a dictionary). The separator ||| should appear once to divide the template into prompt and output. Generally, the prompt should provide information on the desired behavior, e.g., text passage and instructions, and the output should be a desired response.

Here's an example for AG News:

{{text}}
Is this a piece of news regarding world politics, sports, business, or technology? |||
{{ ["World politics", "Sport", "Business", "Technology"][label] }}

Contributing

This is very much a work in progress, and help is needed and appreciated. Anyone wishing to contribute code can contact Steve Bach for commit access, or submit PRs from forks. Some particular places you could start:

1. Try to express things! Explore a dataset and tell us what's hard to do to create templates you want
2. Look in the literature. Are there prompt creation methods that do/do not fit well right now?
3. Scalability testing. Streamlit is lightweight, and we're reading and writing all prompts on refresh.

See also the design doc.

Before submitting a PR or pushing a new commit, please run style formattings and quality checks so that your newly added file look nice:

make style
make quality

Known Issues

Warning or Error about Darwin on OS X: Try downgrading PyArrow to 3.0.0.