PromptSource is a toolkit for creating, sharing, and using natural language prompts, born from the BigScience research workshop. It treats prompts as first-class, reusable artifacts: functions that map a dataset example to a natural-language input and target, stored in structured files anyone can browse, remix, and apply.
PromptSource is hosted on GitHub, where it has earned more than 3,000 stars under the Apache-2.0 license. Its centerpiece is P3, the Public Pool of Prompts: a growing collection that, as of January 2022, already held roughly 2,000 English prompts covering more than 170 English datasets, the same resource used to train instruction-tuned models such as T0. Prompts are written in Jinja, a simple templating language, and saved as standalone structured files, so a prompt is easy to read, diff, and version. A simple API lets you load the thousands of existing prompts programmatically and apply them to dataset examples, while an included web interface lets you browse the collection and author new prompts interactively. The project emerged from the research moment when GPT-3, FLAN, and T0 showed that prompts and multitask fine-tuning unlock strong zero-shot generalization; PromptSource gave the community shared infrastructure for that work instead of everyone hand-writing prompts in private notebooks. The repository’s last updates were in 2023, but the toolkit and the P3 collection remain a reference resource for prompt engineering and instruction tuning.
PromptSource is free and open source under the Apache-2.0 license, and the P3 prompt collection is publicly available, including through Hugging Face.
PromptSource serves NLP researchers, prompt engineers, and instruction-tuning teams who want a large, curated prompt library and a sane format for managing prompts as versioned artifacts instead of scattered strings in code.