snorkel

Why we included this project

Labeling training data is often the biggest bottleneck in machine learning, and Snorkel takes that problem head on. Instead of paying people to hand-label thousands of examples, you write small, noisy rules: regex patterns, heuristics, or distant supervision signals. Snorkel then models how accurate each rule is and how the rules relate to each other, and combines their votes into probabilistic labels you can feed directly into a downstream classifier. That makes it a good fit for teams where domain experts can express what a good label looks like but don't have weeks to spend annotating, especially in fields like clinical text or legal documents where manual labeling is slow and expensive. If you have unlabeled data and rough heuristics, Snorkel is a proven route to turn that combination into usable training data.