PyThaiNLP/attacut

Built on a 3-layer dilated CNN architecture that processes syllable and character features, AttaCut achieves 91% word-level F1 on the BEST benchmark while running 6x faster than previous state-of-the-art approaches. It integrates with PyTorch and provides both command-line and Python APIs for immediate use, with support for custom model retraining on user datasets. The toolkit includes pre-trained models (`attacut-sc` and `attacut-c`) optimized for different accuracy-speed tradeoffs in Thai NLP pipelines.

94 stars and 4,237 monthly downloads. Used by 1 other package. No commits in the last 6 months. Available on PyPI.