PromptSource is a toolkit for creating, sharing, and using natural language prompts. The Prompt Engineering WG used it to create the largest open-source collection of English prompts: 2,000 prompts for 170 datasets – a massive effort by 50 contributors from 25+ institutions.
A prompt is a snippet of text that specifies an NLP task. Prompts have been an extremely useful method recently in NLP (T0, FLAN, GPT-3, etc.). The lightweight tool allows researchers and engineers to create a large number of prompts for a large number of datasets.
Find PromptSource on GitHub:
https://github.com/bigscience-workshop/promptsource
The PromptSource tooling was used to create P3 – a dataset of 2,000 prompts that T0 was trained on.