Neo4j Cypher Query Generation Dataset

44,387 instances of question-to-Cypher query pairs, curated by Neo4j's GenAI team. The dataset is split into 39,554 training and 4,833 testing examples, each containing a question, schema, and Cypher query triplet. It was published on HuggingFace and updated in August 2025.

Use Cases

• Train a transformer model to generate Cypher queries from natural language questions using the 'question' and 'cypher' fields.
• Fine-tune a language model for schema-aware query generation by leveraging the 'schema' field alongside questions.
• Benchmark the accuracy of text-to-Cypher systems on the held-out test set of 4,833 instances.
• Develop a retrieval-augmented generation system that uses the schema context to produce correct graph queries.
• Study the generalization of query generation models across different database schemas represented in the dataset.

Strengths

• Contains 44,387 total instances for model training and evaluation.
• Provides a clear train/test split with 39,554 and 4,833 instances respectively.
• Each instance includes a structured 'question, schema, cypher' triplet.

Limitations

• Specific source datasets and their original sizes are not detailed, limiting traceability.
• Potential for bias towards the types of queries and schemas present in the aggregated source datasets.
• No information on the diversity or complexity range of the Cypher queries included.

Provenance

Source

Aggregated from multiple publicly available datasets.

Collection Method

Instances were cleaned and organized from public sources into a standardized triplet format.

Freshness

Last updated August 2025.