Math2Vec: Multi-View Mathematical Concept Embeddings

A training dataset for mathematical embedding models built on the principle that concepts can be expressed in multiple surface forms. It includes informal natural language, rephrasings, and Lean 4 type signatures and declarations, designed for contrastive embedding training. The dataset was created by uw-math-ai and was last updated on 2026-05-29.

Use Cases

• Train contrastive embedding models based on the multi-view structure of mathematical concepts.
• Align informal natural language descriptions with formal theorem prover code.
• Develop models for mathematical concept retrieval and similarity based on alternative rephrasings.
• Fine-tune language models for understanding mathematical notation and Lean 4 declarations.

Strengths

• Designed around a specific principle of multi-view representation for mathematical concepts.
• Includes structured configs for different training purposes, such as 'mathlib_views'.
• Integrates formal theorem prover (Lean 4) code with informal natural language descriptions.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

uw-math-ai

Collection Method

Likely constructed from mathematical libraries and theorem prover code, but specific gathering method is not detailed.

Freshness

Last updated 2026-05-29 02:18:48; freshness should be verified.