thanhdath/FINER-SQL-0.5B-BIRD

FINER-SQL-0.5B-BIRD: A Compact Text-to-SQL Model

This model, developed by thanhdath, is a 0.5 billion parameter Text-to-SQL solution, fine-tuned from griffith-bigdata/Qwen-2.5-Coder-0.5B-SQL-Writer. It leverages the GRPO algorithm combined with FINER-SQL dense rewards (Memory + Atomic) to enhance SQL generation capabilities.

Key Capabilities

• High Accuracy on BIRD Dev: Achieves 50.85% execution accuracy on the BIRD Dev benchmark (n=30, value-aware voting), demonstrating strong performance for its size.
• Efficient Resource Usage: Designed to run on GPUs with 4-8 GB of memory, making it suitable for deployment-friendly environments.
• Robust SQL Generation: Excels at converting natural language questions and database schemas into correct SQL queries, supporting complex scenarios.
• Scalability of Rewards: Demonstrates that GRPO + FINER rewards effectively scale down to smaller model sizes while retaining significant performance gains.
• Inference Pipeline: Recommends a robust evaluation pipeline involving generating multiple candidates (n=30), executing them, and applying value-aware voting (vav) for optimal results.

Good For

• Text-to-SQL Applications: Ideal for systems requiring accurate SQL query generation from natural language inputs.
• Resource-Constrained Environments: Its small parameter count and efficient memory footprint make it suitable for deployment where larger models are impractical.
• BIRD Benchmark Tasks: Specifically fine-tuned and evaluated on the BIRD dataset, making it a strong candidate for similar complex database querying tasks.
• Research in Reward-Based Fine-tuning: Showcases the effectiveness of GRPO and FINER-SQL rewards in improving model performance for code generation.