This write-up documents a training experiment pushing a 0.9-million-parameter language model to a token-per-parameter ratio of roughly 222,000:1 — far beyond the Chinchilla-optimal ratio of about 20:1. Performance peaked around 20 billion training tokens (a 22,000:1 ratio) before declining monotonically by 27.3% over the remaining 160 billion tokens of training. The results suggest that for tiny language models, the useful token-per-parameter range tops out around 22,000–30,000:1, with training beyond that range actively degrading the model rather than plateauing.
