Small-Text addresses the challenge of efficiently labeling training data for text classification, particularly when labeled data is scarce. It offers state-of-the-art active learning strategies, allowing users to easily combine pre-implemented query strategies, initialization methods, and stopping criteria with classifiers from scikit-learn, PyTorch, or Hugging Face Transformers. This accelerates the development of supervised text classification models by intelligently selecting the most informative data points for manual annotation, benefiting researchers and practitioners alike.

How It Works

The library provides a unified interface for active learning workflows. Users can select from various scientifically evaluated components and integrate them with popular machine learning frameworks. It supports GPU acceleration via PyTorch and seamless integration with Transformers for leveraging advanced text classification models. This modular design facilitates experimentation and application building, optimizing the data labeling process by reducing manual annotation effort.

Quick Start & Requirements

• Installation: pip install small-text for a slim installation; pip install small-text[transformers] for full installation including transformer support.
• Prerequisites: Python 3.9 or newer. CUDA 10.1 or newer is required for GPU acceleration.
• Examples: Several example notebooks are provided, covering binary classification, PyTorch multi-class classification, transformer-based classification, and active learning with stopping criteria. These can be accessed via links within the project's repository, such as:
  • Intro: Active Learning for Text Classification with Small-Text
  • Stopping Criteria: Using Stopping Criteria for Active Learning
  • SetFit: Active Learning using SetFit
  • Zero Shot: Using SetFit's Zero Shot Capabilities for Cold Start Initialization
  • Argilla: Active Learning via the Argilla User Interface

Highlighted Details

• Unified interfaces for mixing and matching query strategies, classifiers (sklearn, PyTorch, Transformers), initialization strategies, and stopping criteria.
• GPU support via PyTorch; CPU-only use has minimal dependencies.
• Version 2.0.0 (alpha) introduces refined interfaces, new query strategies, improved classifiers, and vector indices.
• Awarded EACL Best System Demonstration for its introductory paper.

Maintenance & Community

Developed by Christopher Schröder at Leipzig University's NLP group (Webis). The project is funded by the Development Bank of Saxony. Contributions are welcomed. A community survey on active learning in NLP was conducted in March 2026.

Licensing & Compatibility

Licensed under the MIT License, which is permissive for commercial use and integration into closed-source projects.

Limitations & Caveats

Version 2.0.0.dev3 is an alpha release and may not have stable interfaces. The project emphasizes its progress and feature set, noting that simple counts do not fully represent its capabilities.