WangNet – 1.8 MB, zero-dependency Numberwang adjudication in 11 languages (github.com)

WangNet, a compact neural network model weighing in at just 1.8 MB, has been released to adjudicate whether a given input qualifies as "Numberwang," a humorous concept from British television. This model is notable for its simplicity; it functions with minimal dependencies, requiring only Python 3.8 or newer and about 100 lines of pure Python code. Users can evaluate numbers—including expressions in eleven languages—via straightforward commands, leveraging the model's ability to discern the “wangness” of numbers with impressive confidence levels. The significance of WangNet lies in its unique approach to language processing without traditional preprocessing steps like tokenization or normalization. It directly handles characters and utilizes a convolutional neural network with 80,804 parameters to achieve high accuracy on its adjudications, reaching an 88.9% success rate. However, the model struggles with arithmetic involving unseen operands, indicating areas for improvement. By employing this model, the AI/ML community can explore lightweight solutions to language processing tasks that embrace creativity and humor while testing their robustness across various languages and number formats.