Snowflake/Qwen-2.5-coder-Arctic-ExCoT-32B: Execution-Guided Text2SQL

Snowflake/Qwen-2.5-coder-Arctic-ExCoT-32B is a 32.8 billion parameter model developed by Snowflake, specifically designed for Text-to-SQL tasks. It is built upon the Qwen-2.5-coder architecture and incorporates the novel ExCoT (Execution-Guided Chain-of-Thought) framework. This framework distinguishes itself by optimizing the model through SQL execution-based DPO, using the actual execution results of generated SQL queries as a feedback signal, rather than relying on expensive human annotations.

Key Capabilities & Performance

• Execution-Guided Optimization: Leverages SQL execution outcomes for direct feedback, enabling scalable and high-quality model optimization.
• State-of-the-Art Text2SQL: Achieved strong results on the BIRD-test benchmark, demonstrating significant improvements in execution accuracy compared to base models and outperforming many general-purpose frontier models.
• Public Dataset Training: Optimized using only public datasets like BIRD and Spider, without requiring additional Text2SQL data.
• High Accuracy: On the BIRD-dev set, this model achieved 68.25% execution accuracy, showcasing its proficiency in complex SQL generation.

Ideal Use Cases

• Complex Database Interaction: Excellent for applications requiring precise conversion of natural language questions into SQL queries for intricate databases.
• Automated Data Analysis: Suitable for tools that automate data retrieval and analysis by generating accurate SQL from user prompts.
• Benchmarking Text2SQL Solutions: Provides a strong baseline and competitive performance for evaluating Text2SQL systems.