I don't think that this produces correct results, as it seems to determine whether each individual number is Numberwang. However, whether a number is Numberwang can also depend on the previous numbers in the sequence or other factors.

So I think as it stands right now this model can only determine whether a number would be Numberwang as the first number of a sequence, but even for that I still wouldn't rely on this in the actual game show.

For those confused by this, here is the Mitchell and Webb "Numberwang" sketch

https://youtu.be/0obMRztklqU?is=qufDLUkqofpr-K1Y

You're playing with fire here, don't forget the tragedy of Colosson. Computers weren't meant to do this.

Amazing, I passed by Mornington Crescent only yesterday!

What's the training corpus, just the canonical Mitchell and Webb sketch? Or ancilliary data like comments and references, or worse synthetic data generated on the primary sources?

Hey! I was not consulted?

How did you all first learn about the original Numberwang? Netflix?

This is hilarious and I miss numberwang. Also, the history of numberwang. And everything to do with numberwang. It was the most prescient sport back in the day.

I'm glad that someone has stepped up to revive the legacy of numberwang.

My friends, this is the single most hilarious thread I have ever experienced on the Interwebs. I am deeply grateful to have found this. I feel like my life is now complete.

what's going on with python that this is considered zero-dependency? requirements.txt, import gradio?

Is this using the rules from the official board game?

Hello I have a question, and my question is what the fuck

Any plans on implementing Wordwang?

Looking forward for the wordwang version.

thats Numberwang!

Numberwang??

Try it in the browser (the forward pass is plain JS, nothing leaves the page): https://huggingface.co/spaces/graafhenk/numberwang-demo

What it is: a character-level CNN with 80,804 parameters. The weights are a 1.79 MB JSON file and inference is about 100 lines of Python standard library — no PyTorch, no NumPy. It runs on a Pi Zero. It accepts digits, number words in eleven languages, arithmetic ("96 divided by 2", "deux fois trois"), Roman numerals, ordinals, clock times, currency, and fictional numbers ("shinty-six"). Anything with no numeric content is correctly ruled out as never able to be Numberwang. Whatever comes to 1 or 44 is Wangernumb and you rotate the board.

Held-out accuracy is 88.9% on 486 probes reserved from training by construction. The ceiling is ~98%, because roughly 2% of training labels are inverted at compilation time, in accordance with long-standing adjudication practice.

For comparison I ran Qwen3-1.7B on the same suite with the four verdicts as a constrained multiple choice: 51.9%, which is 2.3 points above answering "Numberwang" to everything. It answers "Numberwang" to 93% of inputs and never once identifies a Wangernumb. So the accuracy table has a verdict-distribution column, since one number can't tell a model that decides from one that agrees.

Honest weak spot: arithmetic is memorised, not computed. A conv net can't add. On operands reserved from training it gets 60% on symbolic expressions and 44% on foreign-language ones.

Dataset (185k adjudicated utterances), training script, evaluation harness and benchmark are all in the repo and reproduce from a fixed seed. Model card on HF: https://huggingface.co/graafhenk/numberwang