webis-de/small-text

small‑text – Active Learning for Text Classification

What it is

• A Python library that implements active learning workflows for text classification. It lets you start with very few labeled examples, train a classifier, query the most informative unlabeled documents, label them, and repeat until a stopping condition is met.

Key capabilities

Capability	Details
Query strategies	19 built‑in methods (e.g., uncertainty, diversity, margin, entropy, SetFit‑based) that decide which unlabeled texts to ask a human annotator to label.
Initialization	Several ways to bootstrap the first labeled set, including random, clustering‑based, and zero‑shot SetFit initialization.
Stopping criteria	5 ready‑to‑use criteria (budget, performance plateau, confidence thresholds, etc.) to automatically end the active‑learning loop.
Classifier back‑ends	Works with any scikit‑learn estimator, PyTorch models, and Hugging Face transformers. GPU support is optional; a lightweight CPU‑only install is also available.
Extensibility	Interfaces are deliberately modular so you can mix any query strategy, classifier, and stopping rule.

Typical workflow

1. Install – pip install small-text (add [transformers] extra for transformer models).
2. Prepare data – supply a pool of unlabeled texts and (optionally) a small seed set of labels.
3. Choose components – pick a classifier (e.g., sklearn.LinearSVC, a torch.nn.Module, or a transformers model), a query strategy, and a stopping criterion.
4. Run the loop – the library handles training, querying, and updating the labeled set automatically.
5. Export – retrieve the final trained model or the labeled dataset for downstream use.

Why it matters

• Label efficiency – reduces the amount of manual annotation needed for high‑quality text classifiers, which is especially valuable in domains where labeling is costly (biomedicine, social science, legal text, etc.).
• Research‑ready – the library is used in several peer‑reviewed studies and includes reproducibility notes, making it suitable for academic experiments.
• Integration friendly – can be combined with UI tools like Argilla for a full annotation interface, or embedded in custom pipelines.

Getting started

• The repository ships example scripts for binary classification, PyTorch multi‑class, and transformer‑based multi‑class setups.
• Five Jupyter notebooks (available on GitHub and launchable in Google Colab) walk you through an end‑to‑end active‑learning session, stopping‑criteria usage, SetFit integration, zero‑shot cold‑start, and a UI demo with Argilla.

Documentation & community

• Full docs at https://small-text.readthedocs.io covering installation, component reference, reproducibility notes, and a showcase of real‑world applications.
• The project is actively maintained (MIT‑licensed), has CI badges, a DOI for citation, and welcomes contributions via a CONTRIBUTING guide.

When to use it

• You have a large corpus of texts but only a handful of labeled examples.
• You want to experiment with different active‑learning strategies without writing boiler‑plate code.
• You need a library that works with both classic ML models and modern transformer models.

---

small‑text is therefore a genuine, actively maintained software project focused on a core machine‑learning technique (active learning) for text data, fitting squarely within the AI/ML ecosystem.