Is Even Core

This repository contains the minimal public code needed to generate the dataset, train the model locally on CPU, and benchmark saved checkpoints.

Included:

• training (is_even/train.py)
• data generation (is_even/data.py)
• benchmarking (is_even/benchmark.py)
• required support modules (constants.py, modeling.py, tokenizer.py, __init__.py)

Not included:

• model weights
• internal research tooling
• prediction CLI convenience code

Published weights and the model card live at https://hf.co/SnurfyAI/is-even.

Requirements

• Python 3.12
• a local environment that can install torch and transformers

Install

From the repository root:

python -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install .

This installs two console commands:

• is-even-train
• is-even-benchmark

If you prefer not to install the console scripts, you can run the modules directly:

python -m is_even.train
python -m is_even.benchmark --model-dir artifacts/is-even

Train a model

The default training command is:

is-even-train

What it does:

• creates data/is_even_chatml.jsonl if it does not already exist
• trains a CPU model on the generated train split
• evaluates on the validation split during training
• writes the trained model, tokenizer, and training_summary.json to artifacts/is-even

Useful overrides:

is-even-train --dataset-path data/is_even_chatml.jsonl --output-dir artifacts/is-even --epochs 60 --batch-size 64 --learning-rate 1e-3

The main knobs are:

• --epochs
• --batch-size
• --learning-rate
• --train-samples-per-last-digit
• --validation-samples-per-last-digit
• --layers
• --heads
• --embedding-size

Benchmark a checkpoint

Once you have a model directory, run:

is-even-benchmark --model-dir artifacts/is-even --examples 1000

The benchmark prints:

• total examples evaluated
• accuracy
• examples per second
• up to five incorrect predictions

Use the published weights

This repository does not include weights. To benchmark the published model, download the files from https://hf.co/SnurfyAI/is-even into a local directory and point --model-dir at that directory:

is-even-benchmark --model-dir path/to/downloaded-is-even-model --examples 1000

Use the helpers from Python

You can also import the package directly:

from is_even.data import ensure_dataset, number_is_even
from is_even.tokenizer import IsEvenTokenizer

print(number_is_even("42"))

summary = ensure_dataset("data/is_even_chatml.jsonl", IsEvenTokenizer())
print(summary.total_examples)