DeepSeek-Prover V1: 10K-100K Synthetic Lean Mathematical Proofs

DeepSeek-Prover V1 contains between 10,000 and 100,000 synthetic mathematical proof records designed for the Lean proof assistant. Developed by deepseek-ai and released in 2024, this dataset facilitates the training and evaluation of large language models in formal mathematical reasoning.

Use Cases

• Fine-tuning large language models to generate valid Lean code for mathematical theorems
• Benchmarking model accuracy in formal verification using synthetic proof structures
• Augmenting training sets for mathematical reasoning where human-annotated Lean data is scarce

Strengths

• 10,000 to 100,000 records
• Specifically formatted for the Lean proof assistant
• Linked to peer-reviewed research (Arxiv 240514333)

Limitations

• Synthetic data may contain logical artifacts not present in human-written proofs
• Restricted to Lean syntax, limiting utility for other proof assistants like Coq
• Specific column schema is not explicitly documented in the metadata

Provenance

Source

deepseek-ai (Arxiv 240514333)

Collection Method

synthetic

Freshness

Last updated September 2024.