Snorkel

Instead of hand-labeling training data, Snorkel users write labeling functions: small Python heuristics, pattern matchers, or distant supervision rules that each vote noisily on a label. A generative label model then estimates the accuracy and correlations of those functions and combines their votes into probabilistic training labels, letting a downstream model learn from far more data than manual annotation could produce. The approach, called weak supervision, originated at Stanford in 2015 and became the foundational programmatic labeling project before spawning the company Snorkel AI. The library installs with pip or conda, runs on CPU, and covers labeling functions, transformation functions for data augmentation, and slicing functions for monitoring important subsets. It is Apache 2.0 licensed with documentation at snorkel.readthedocs.io and sits around 6,000 GitHub stars. The repository is now in maintenance mode, with the core team focused on the commercial Snorkel Flow platform, but it remains the reference implementation and the standard teaching codebase for weak supervision.