small-text

Why we included this project

Labeled data is usually the bottleneck when you train a text classifier, and small-text attacks that problem directly. It implements active learning: instead of labeling a random sample, you let the model flag the unlabeled documents that would teach it the most, then annotate those. The query strategies, initialization routines, and stopping criteria are interchangeable and plug into scikit-learn, PyTorch, and Hugging Face transformers, so the same loop works on a lightweight CPU model or a GPU-backed transformer. That modularity suits both controlled experiments and a real labeling pipeline, and the documentation includes notebooks and worked examples. If you need to bootstrap training data for document classification or topic tagging without a large pre-labeled corpus, this is a concrete way to cut annotation cost.