TOME‑Math turns the text of 2,118 university‑level mathematics textbooks into a structured corpus of about 380,000 proof–supervision records. Each record keeps the chapter sequence, links between theorems and their proofs, references to supporting propositions, and a clear trail back to the original source book.

This structured data was fed into a Qwen3‑30B‑A3B‑Base language model. After one‑epoch training on the full TOME‑Math set, the model’s average accuracy on six held‑out math benchmarks increased from 29.17 % to 45.83 %. That performance surpasses all three published instruction‑data baselines tested in aggregate.

The size and structure of the corpus are shown in a table that lists TOME‑Math’s 380,292 questions and answers, gained from 2,118 titles, alongside other datasets. The table also highlights that the records retain chapter order, reference links, and provenance.