TOME-Math turns 2,118 textbooks into 380K structured proof records
A corpus from Intern Large Models converts 2,118 mathematics textbooks into about 380K proof-supervision records that keep chapter order and proposition-proof links.
Mathematical supervision for language models usually arrives as isolated problem-solution pairs. A paper from Intern Large Models argues that flattening the data this way discards the curricular organization and evidential structure the original sources already carried: the order in which material is introduced, which proof belongs to which proposition, and which earlier results a proof leans on.
What the corpus keeps
TOME-Math, short for Textbook-Oriented Mathematical Evidence, converts 2,118 textbooks spanning the core areas of university-level mathematics subjects into approximately 380K structured proof-supervision records. The conversion is built to preserve the structure the abstract names: chapter order, proposition-proof links, dependency references, and source provenance. A record therefore sits inside a curriculum instead of floating free, and a proof keeps a pointer back to the result it proves and to the material it depends on.
That makes the corpus two things at once. It is training data, and it is a way to vary data organization on purpose: with chapter order and dependency links intact, the question of whether presenting supervision in curricular order helps can be asked rather than assumed.
What the authors report
Using those structural signals, the authors explore curriculum learning across multiple mathematical subjects and characterize scaling patterns with supervision volume and source breadth. At full scale, fine-tuning on TOME-Math raises the unweighted mean score across their held-out evaluations and established public benchmarks by 16.7 percentage points over the base model. The authors also report that this exceeds all three public instruction-data baselines they evaluated, in aggregate.
Both of those are the authors’ own figures from the abstract, not an independent evaluation. The abstract presents them as evidence for two claims: that structured textbooks are a scalable source of mathematical supervision, and that they give an empirical framework for studying how data organization affects mathematical reasoning.
What the abstract leaves open
The abstract does not say which base model was fine-tuned, nor which held-out evaluations and public benchmarks make up the mean score that moved by 16.7 percentage points. The three public instruction-data baselines are not named either, and that comparison is reported in aggregate rather than one benchmark at a time. There is no per-subject breakdown at this level of detail, so how much of the gain comes from curricular ordering as opposed to sheer volume is not something the abstract settles.
How the 2,118 textbooks were obtained, and under what terms, is also left to the full paper. The authors state they will release the corpus to support further research on structured mathematical supervision; the abstract gives no date, and the paper page lists no code repository.
Source: Scaling Mathematical Proof Supervision from Structured Textbooks (arXiv 2610.40394)