WangNet – 1.8 MB, zero-dependency Numberwang adjudication in 11 languages

summary

The author presents a small, self-contained neural network designed to classify whether a given input is 'Numberwang' based on a variety of numeric formats. The model is highly efficient, requiring only a 1.8 MB JSON file for weights and a small amount of pure Python code for inference. By bypassing heavy libraries like PyTorch or NumPy, it demonstrates that a specialized classification task can be achieved with minimal dependencies. The system processes raw characters directly to handle multiple languages and formats without a separate tokenizer.

key points

1. The model uses a convolutional neural network architecture with global max pooling to classify inputs into four distinct categories. It avoids traditional preprocessing by learning character embeddings and patterns directly from the weights.
2. The network supports a wide array of inputs including eleven different languages, Roman numerals, and basic arithmetic expressions. However, it relies on memorization rather than computation, meaning it may struggle with complex or unseen mathematical operations.
3. Performance is measured at 88.9% accuracy across held-out data, which is close to the theoretical ceiling given noisy training labels. The inference engine is designed for portability, running on Python 3.8+ with no external dependencies.

who should read it

Software developers and AI enthusiasts interested in lightweight model deployment and minimal dependency architectures. It is also suitable for fans of the comedy show whose concept the project is based on.

discussion warm