PromptSource
github.com
- Free plan
- Web
- API
- Self-hosted
What PromptSource is
PromptSource is a free, open-source toolkit for developers creating, sharing and using natural-language prompts. Its web interface lets users write Jinja templates and inspect outputs against examples. The application includes Sourcing, Prompted dataset viewer and Helicopter view modes. Prompts are stored in standalone structured files, and the API provides Template, Metadata, DatasetTemplates and TemplateCollection classes for managing prompts and metadata. Prompts can be applied to examples from the Hugging Face Datasets library. The repository describes the P3 collection as about 2,000 English prompts for more than 170 English datasets as of January 20, 2022. Metadata can record original-task status, answer-choice presence and evaluation metrics. Users can install it with pip and launch the local app with Streamlit. Hosted versions support browsing, although the BigScience hosted app disables Sourcing mode. Some datasets require manual downloads. The project uses Apache-2.0.
Who it suits
PromptSource suits developers and contributors building or evaluating prompt templates with Python and Hugging Face datasets.
Good to know, and what to watch
- Free Apache-2.0 open-source project
- Jinja prompt editor with example outputs
- API classes for templates and metadata
- Prompts integrate with Hugging Face Datasets
- Supports local Streamlit deployment
- Stable prompt creation currently requires Python 3.7
- Some datasets need manual downloads
- Hosted BigScience app disables Sourcing mode
- Manual setup can be required
The verdict
PromptSource offers a developer-focused prompt workspace with both a web app and Python API. It is useful for teams that want structured templates, dataset examples and local control.
What the maker publishes
- API
- The API provides Template, Metadata, DatasetTemplates, and TemplateCollection classes to store, manipulate, and use prompts and metadata. source
- App modes
- The app has Sourcing, Prompted dataset viewer, and Helicopter view modes. source
- Dataset integration
- Prompts can be applied to examples from the Hugging Face Datasets library. source
- Hosted access
- The README links to hosted versions for browsing existing prompts and notes that the BigScience hosted app disables Sourcing mode. source
- Hosted version
- The README links to a hosted version for browsing existing prompts and says its Sourcing mode is disabled. source
- Installation
- The README gives pip install promptsource as an installation command for users who do not intend to create new prompts. source
- Intended audience
- The package metadata lists developers as the intended audience. source
- Intended contributors
- The contribution guide says a way to grow P3 is to write prompts for new datasets and submit them in a pull request. source
- License
- The public repository is licensed under Apache-2.0. source
- Local app
- The README says to launch the app locally with streamlit run promptsource/app.py after completing setup. source
- Local install
- The README gives pip install promptsource as an installation option and documents launching the app locally with Streamlit. source
- Manual datasets
- Some datasets require manual download into ~/.cache/promptsource by default, and the location can be overridden with PROMPTSOURCE_MANUAL_DATASET_DIR. source
- Manual dataset setup
- Some datasets require manual download into ~/.cache/promptsource by default, and the path can be changed with PROMPTSOURCE_MANUAL_DATASET_DIR. source
- Prompt collection
- The README describes a collection of about 2,000 English prompts for more than 170 English datasets as of January 20, 2022. source
- Prompt editor
- Its web-based GUI lets developers write prompts in a templating language and view outputs on examples. source
- Prompt format
- Prompts are stored in standalone structured files and written in Jinja. source
- Prompt library
- The repository describes a collection called P3, with about 2,000 English prompts for more than 170 English datasets as of January 20, 2022. source
- Prompt metadata
- Prompt metadata can include the original task flag, whether answer choices appear in the prompt, and evaluation metrics. source
- Purpose
- PromptSource is a toolkit for creating, sharing, and using natural language prompts. source
- Runtime requirement
- The package metadata specifies Python >=3.7,<4.0, and the README says creating prompts through the interface currently requires Python 3.7 for stability. source
Checked against the maker's pages on 9 Oct 2026.