bio2C v1

Benchmark

A biomedical benchmark containing 2,500 manually curated natural language questions and corresponding Cypher queries over miRNA-KG.

Statistics

Property	Value
Version	v1
Domains	Biomedical
Languages	English
Curation	Manually curated
Total NL–Cypher pairs	2,500
Graph-executable pairs	2,500
Complexity levels	5
Complexity definition	The dataset is divided into five categories: node-level, 1-hop, 2-hop, 3-hop, and advanced. Each category contains 500 pairs.

Database: miRNA-KG T2C

Endpoint

Property	Value
Web endpoint	https://neo4j.biodata.di.unimi.it
Bolt endpoint	neo4j+s://helix.biodata.di.unimi.it:7687
Username	Text2Cypher
Password	Text2Cypher
Database	mirnakgt2c
Neo4j dump	Download dump

Best reported results

• Scientific source: DOI: 10.21203/rs.3.rs-10463153/v1

Metric	Score	Model	Technique(s)
Jaro–Winkler	94.7%	GPT-4oFT	RAG, RAG+O, S+RAG
Jaccard	86.4%	GPT-4oFT	RAG
Coverage	84.3%	GPT-4oFT	RAG
Pass@1	89.3%	GPT-4oFT	RAG, S+RAG+O

Reference and license

• License: TBA
• Associated article: DOI: 10.21203/rs.3.rs-10463153/v1
• Last reviewed: 2026-08-03

Notes

• TBA