This page shows all available databases in the BIRD benchmark that can be used for generating incorrect SQL queries.
The BIRD benchmark is a comprehensive collection of databases with natural language questions and their corresponding SQL queries. It's designed for evaluating text-to-SQL systems across diverse schemas.
The datasets contain real-world database schemas with varying complexity levels, making them ideal for generating realistic SQL error patterns across different domains.
The following database schemas have SQLite files available for testing and validation. Only queries from these databases will produce fully validated results:
Database with information about California schools, test scores, and meal programs.
+ 86 more queries in this database
Banking database with accounts, clients, loans, and transactions.
+ 103 more queries in this database
Database for chemical compounds with toxicology/carcinogenicity data.
+ 142 more queries in this database
Database for card games.
+ 188 more queries in this database
Database for codebase community.
+ 183 more queries in this database
Database containing superhero information including powers, attributes, and publishers.
+ 126 more queries in this database
Database for formula 1.
+ 171 more queries in this database
Database for european football 2.
+ 126 more queries in this database
Database for thrombosis prediction.
+ 160 more queries in this database
Database for university student clubs, members, and activities.
+ 155 more queries in this database
Database for debit card specializing.
+ 61 more queries in this database