Generate Realistic but Incorrect SQL Queries

Expanded Database Coverage!

We now support multiple database schemas from the BIRD benchmark with hundreds of additional seed queries!

Available SQLite databases: Superhero, California Schools, Toxicology, Student Club

Explore All Databases Explore Error Taxonomy
Generation Parameters
How many random natural language queries to select (max 100)
Maximum number of incorrect SQL queries to generate for each natural language query
Seed for random selection (for reproducibility)
Filter which database types to include in the random selection
About

This tool generates realistic but incorrect SQL queries based on natural language questions from the BIRD benchmark.

The tool uses GPT-4o to generate queries that represent common mistakes a human might make when writing SQL, with full support for multiple database schemas.

Features:
  • Generates syntactically valid but semantically incorrect SQL queries
  • Validates that incorrect queries produce different results from correct queries
  • Analyzes error patterns and categorizes common mistakes
  • Exports results in JSONL or JSON format