Distil-Qwen3-0.6B-Text2SQL Overview
This model, developed by Distil Labs, is a compact 0.6 billion parameter Qwen3-based model specifically fine-tuned for converting natural language questions into SQL queries. It leverages knowledge distillation from the much larger DeepSeek-V3 model, allowing it to achieve high performance at a significantly smaller size.
Key Capabilities & Performance
- Text-to-SQL Conversion: Excels at transforming natural language questions into accurate SQL queries.
- High Accuracy: Achieves 74% on LLM-as-a-Judge accuracy, demonstrating a 2x improvement over its base Qwen3-0.6B model and approaching the performance of the 685B parameter DeepSeek-V3 teacher model.
- Efficiency: With only 0.6 billion parameters and a 40,960 token context length, it is designed for lightweight and fast inference, suitable for local and edge deployments.
- SQL Feature Support: Handles a range of SQL complexities including SELECT, WHERE, COUNT, SUM, AVG, MAX, MIN, JOIN, GROUP BY, HAVING, ORDER BY, LIMIT, subqueries, and UNION.
- Training: Fine-tuned on approximately 10,000 synthetic examples generated by DeepSeek-V3, ensuring robust performance.
Ideal Use Cases
- Natural Language Interfaces: Building conversational interfaces for databases.
- SQL Query Assistance: Providing automated SQL generation and autocompletion.
- Database Chatbots: Powering chatbots for business intelligence and data exploration.
- Educational Tools: Assisting in learning and practicing SQL.
- Edge Deployment: Suitable for applications requiring on-device or resource-constrained inference due to its compact size.