Make a small open model better at math

No riders yet

Make a 1B model reason better.

Improve a small open-weight language model (around 1B parameters) on grade-school and competition math with a reproducible, affordable recipe. Establish baselines on public benchmarks with fixed prompts, test one training or inference-time idea at a time, check for test-set contamination, and report accuracy with confidence intervals, cost and full scripts.