QData/TextAttack

Implements modular attack "recipes" combining transformations (word swaps, BERT masking), constraints (semantic similarity, POS consistency), and search methods (genetic algorithms, greedy-WIR) to generate adversarial examples across classification, entailment, and sequence labeling tasks. Provides pre-built integrations with popular models (BERT, DistilBERT, LSTM) and datasets, with parallel GPU support and command-line access to attack execution, dataset augmentation, and model training workflows.

3,377 stars and 7,388 monthly downloads. No commits in the last 6 months. Available on PyPI.