Prodigy

Prodigy is a proprietary annotation tool that runs on your own machines, including air-gapped systems without an internet connection. It's for developers and research teams building training and evaluation datasets for custom AI models. The Python library includes a web application where annotators can label data without programming knowledge.

Its tasks cover named entity recognition, text classification, object detection and image segmentation, as well as model evaluation. Teams can define labels through real examples, review model suggestions and train models from the resulting annotations. Built-in workflows provide a starting point for common tasks; custom Python workflows let developers adapt the data feeds, processing and annotation interface to their own projects.

Prodigy integrates closely with spaCy and has plugins for using and training Hugging Face models. Teams can also use the annotations to train models with PyTorch or TensorFlow. LLM API integrations support model-assisted labeling, including workflows where people correct generated annotations. Those external API calls use a cloud service; the annotation application itself runs locally and doesn't phone home.

The tool can run on premises or on a cloud provider you choose. Local workflows keep data and models under your control, and you own the models you produce. Developers can combine interface components or build custom interfaces with HTML, CSS and JavaScript.