Nemotron-Math-Proofs-v1: 580,000 Proofs and Lean 4 Formalizations

NVIDIA released Nemotron-Math-Proofs-v1 in early 2025, providing a large-scale collection of approximately 580,000 natural language proof problems. The dataset includes 550,000 formalizations into Lean 4 theorem statements and 900,000 model-generated reasoning traces for mathematical distillation.

Use Cases

• Training language models to translate natural language proofs into Lean 4 formal statements
• Distilling mathematical reasoning capabilities using the 900,000 reasoning traces
• Fine-tuning models for automated theorem proving and formal verification

Strengths

• Contains 550,000 formalizations in Lean 4 for automated theorem proving
• Includes 900,000 model-generated reasoning traces for distillation
• Large-scale coverage of 580,000 natural language proof problems

Limitations

• Synthetic reasoning traces may contain logical errors or hallucinations common in model-generated data
• Scope is limited to mathematical proofs and Lean 4 formalization

Provenance

Source

NVIDIA (Arxiv 2508.10925, 2508.03613, 2512.15489)

Collection Method

Synthetic data generation and distillation from multi-mode supervision

Freshness

Last updated January 2026, reflecting state-of-the-art synthetic reasoning as of late 2024.