thanhdath/FINER-SQL-0.5B-Spider

FINER-SQL-0.5B-Spider: A Specialized Text-to-SQL Model

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

Key Capabilities

• High Accuracy for its Size: Achieves 75.0% Execution Accuracy on the Spider Dev dataset (n=30, value-aware voting), making it a highly effective small-scale Text-to-SQL model.
• Efficient Resource Usage: Designed to run efficiently on GPUs with 4-8 GB of memory.
• Specialized Fine-tuning: Demonstrates that dataset-specific specialization significantly improves performance, outperforming the 0.5B BIRD model by 6.4 percentage points on Spider Dev.
• Robust Inference Pipeline: Recommends a pipeline involving generating 30 candidates with temperature 1.0, executing them, and applying value-aware voting (vav) for optimal results.

Good For

• Text-to-SQL Applications: Ideal for converting natural language questions into executable SQL queries, particularly for SQLite databases.
• Resource-Constrained Environments: Suitable for deployment where computational resources are limited, thanks to its small parameter count.
• Spider Dataset Tasks: Specifically optimized for tasks related to the Spider dataset, showing strong performance across various query hardness levels.
• Developers needing SQL generation: Provides a reliable and accurate tool for automating SQL query creation from user input.