Scaling Mathematical Proof Supervision from Structured Textbooks
Fine-tuning on TOME-Math lifted the mean evaluation score by 16.7 percentage points over the base model.
The corpus converts 2,118 university mathematics textbooks into roughly 380,000 structured proof-supervision records. It preserves chapter order, proposition-proof links, dependency references and source provenance that isolated problem-solution pairs discard. At full scale, it also outperformed all three public instruction-data baselines in aggregate across the authors’ held-out evaluations and established public benchmarks. The authors plan to release the corpus.
HF Daily Papers' note
The corpus converts 2,118 university mathematics textbooks into roughly 380,000 structured proof-supervision records. It preserves chapter order, proposition-proof links, dependency references and source provenance that isolated problem-solution pairs discard. At full scale, it also outperformed all three public instruction-data baselines in aggregate across the authors’ held-out evaluations and established public benchmarks. The authors plan to release the corpus.
HF Daily Papers' note
score 5