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

WangNet, a compact neural network, determines whether a number is deemed Numberwang in 11 languages. The entire model is a mere 1.8 MB JSON file, accompanied by just 100 lines of pure Python inference code, requiring no additional installations like PyTorch or NumPy. To access the model, users can clone it and run it without any arguments for an interactive session, which necessitates Python 3.8 or above.

The crucial aspect of Numberwang is that it resides within the number itself, not the language in which it's expressed. For instance, "four," "vier," "quatre," and "cuatro" all share the same verdict. The network directly processes characters, without the need for tokenization, normalization, or rule engines during inference. Digits, operators, and verdicts are all encoded in the weights, with model.json housing the entirety.

A hosted version of WangNet is available on Hugging Face Spaces. For a local demonstration, only Gradio is required, although the model itself doesn't necessitate it. The network achieved an impressive 88.9% macro-F1 score over 486 held-out adjudications, falling slightly short of the theoretical 98% ceiling. This discrepancy arises because approximately 2% of training labels were inverted, following established adjudication practices.

The network's primary weakness lies in handling arithmetic operations on unseen operands, with accuracy ranging from 44% to 72%. While small, common expressions like "5*2" yield reliable results, larger calculations such as "904 * 3" are subject to educated guesses. It's essential to note that the network memorizes rather than computes, meaning that well-known arithmetic expressions are handled accurately while others are left to the network's memorization capabilities.

Users should exercise caution regarding the model's warranty, as it explicitly states that no express or implied guarantee is provided regarding any particular number's Numberwang status. In essence, WangNet is a small neural network that decides whether a number is Numberwang, with no additional dependencies or prerequisites.