| ---
|
| library_name: transformers
|
| pipeline_tag: text-classification
|
| tags:
|
| - transformers
|
| - pytorch
|
| - text-classification
|
| - synthetic-data
|
| - parity
|
| license: mit
|
| ---
|
|
|
| # Is Even
|
|
|
| `Is Even` is a tiny decoder-only transformer trained on chat-formatted synthetic data from SnurfyAI that predicts whether a decimal input is even.
|
|
|
| This repository contains the published model weights and metadata for the released checkpoint.
|
|
|
| ## What The Model Does
|
|
|
| - Input: a decimal numeral rendered inside a fixed ChatML prompt
|
| - Output: `true` for even numbers and `false` for odd numbers
|
| - Tokenization: strict digit-level tokenization with a tiny fixed vocabulary
|
|
|
| The task is mathematically determined by the final decimal digit, but the model was trained as a small transformer classifier on synthetic examples.
|
|
|
| ## Prompt Format
|
|
|
| ```text
|
| <|im_start|>user
|
| 123456<|im_end|>
|
| <|im_start|>assistant
|
| ```
|
|
|
| The expected completion label is `true` or `false`.
|
|
|
| ## Training Code
|
|
|
| You can train the model yourself using the official SnurfyAI training codebase:
|
|
|
| - [SnurfyAI Is Even Core](https://github.com/SnurfyAI/is-even)
|
|
|
| The repository contains the code needed to generate the dataset, train the model, and benchmark saved checkpoints.
|
|
|
|
|
| ## Research Paper
|
|
|
| - [A Microscopic Study of Decimal Parity in a Tiny Causal Transformer](https://ai.snurfy.net/research/is-even)
|
|
|
| The accompanying paper analyzes the model’s generalization behavior, attention structure, and causal dependence on the final digit.
|
|
|
