Training a small language model on the KJV

i decided to train a small language model from scratch to make sure i actually understand how they work. so i asked Anthropic's Fable 5/Opus 5 to create a small model and explain each stage. we ran the code, changed the settings, chose the comparisons and examined the results. OpenAI's Sol 5.6 later assessed and revised the text, with look-ins from GLM 5.3 and a couple others. where the article reports a measurement, we ran it; where it extrapolates, we say so.

it's a 10M model trained on the kjv. a single training run on my macbook pro lasts about 19 minutes.

here's how it went.