trl-lab/qwen3.5-2b-grpo-bird

M_BIRD: A Specialized Text-to-SQL Agent

M_BIRD is a 2.3 billion parameter text-to-SQL agent model developed by trl-lab, part of the SQaLe model family. It is built upon the Qwen3.5-2B architecture and trained using the GRPO (Generalized Reinforcement Learning with Policy Optimization) method on the BIRD training split. Unlike traditional text-to-SQL models, M_BIRD operates as an agent, exploring the database via tools, testing queries, and submitting them without the schema being part of its initial prompt.

Key Capabilities

• High Accuracy on BIRD: Achieves 54.7% on the BIRD dev set, representing a +35.4 point improvement over the untrained base model.
• Reinforcement Learning Only: Trained from a base checkpoint without supervised warm-start or reasoning traces from larger teachers.
• Agentic Database Interaction: Utilizes a suite of tools (e.g., list_tables, describe_table, run_query) to interact with and understand the database schema dynamically.
• Robust Query Generation: Capable of generating SQL queries for complex natural language questions, even when schema information is withheld.

Good For

• Text-to-SQL Applications: Ideal for scenarios requiring conversion of natural language questions into SQL queries, particularly for SQLite databases.
• Agentic Systems: Suitable for integration into agent-based systems where models need to interact with external tools and environments.
• Research in Small, Specialized LLMs: Serves as a reference model for understanding the impact of training data on small, specialized text-to-SQL models, especially when compared to its siblings M_SQaLe and M_SynSQL.
• Environments with Withheld Schemas: Excels in settings where the database schema is not directly provided in the prompt, relying instead on tool-based exploration.