// the take
small-text is a Python library for active learning in text classification: strategies for picking which unlabeled examples to annotate next so you reach good accuracy with less labeled data. It works across sklearn, PyTorch, and Hugging Face transformers/SetFit classifiers, and is aimed at ML researchers and engineers building classifiers where labeling budget is the bottleneck, not compute.
The backend abstraction is real, not cosmetic — you can swap an SVM for a transformer classifier and keep the same query strategy and active learner loop. It ships 19 query strategies and 5 stopping criteria that are actually implemented and tested, rather than a paper's worth of pseudocode. Test suite includes unit, integration, and approval tests with codecov wired up, and the slim pip install keeps sklearn-only users off the PyTorch/transformers dependency tree. It also has a real publication (EACL 2023, won best demo) behind the design, which is a decent signal the abstractions were thought through rather than bolted on.
The version worth installing, 2.0.0.dev4, is explicitly alpha with no interface stability guarantee, while the last actual stable release (1.4.1) is from August 2024 — you're choosing between a stale stable API or an unstable one that can break under you before 2.0 ships. It's a one-maintainer academic project (Christopher Schröder, Leipzig/Webis) tagged 'looking-for-contributors,' so bus-factor and long-term maintenance are real risks outside grant-funded attention. Active learning only pays off if you have a labeling loop to plug it into — this library gives you the selection algorithm, not an annotation UI, so you're on your own for that half (there's an Argilla showcase notebook, but it's a third-party integration, not a built-in path). It's also built for experimentation and offline research loops, not a production serving story — no guidance on running this as part of a live labeling pipeline at scale.
We read the README, the directory tree and a key source file in November 2026. Nobody pays to be listed. How we judge →