AlioLeuchtmann/qwen-3.5-sql-27B-Distill-2B
AlioLeuchtmann/qwen-3.5-sql-27B-Distill-2B is a 2.3 billion parameter text-to-SQL model distilled from a Qwen 27B teacher, specifically fine-tuned for SQLite. This model achieves 68.47% execution accuracy on the official BIRD benchmark, matching 70B-class systems despite its significantly smaller size. It excels at converting natural language questions into SQL queries, making it highly efficient for specialized database interaction tasks.
Loading preview...
Model Overview
AlioLeuchtmann/qwen-3.5-sql-27B-Distill-2B is a compact yet powerful 2.3 billion parameter model specialized in text-to-SQL conversion for SQLite databases. It was developed by Alio Leuchtmann through execution-verified structured chain-of-thought distillation from a Qwen 27B teacher model, without the use of reinforcement learning or extensive test-time reasoning.
Key Capabilities & Performance
- Exceptional Efficiency: Achieves 68.47% Execution Accuracy (EX) on the official BIRD benchmark, performing comparably to 70B-class models and outperforming many larger general-purpose LLMs like Claude 4.5 Sonnet and GPT-4o-based pipelines in its specialized domain.
- Distillation Success: Demonstrates that knowledge distillation can compress significant task competence into a much smaller model, enabling deployment on consumer GPUs or edge hardware.
- Structured Analysis: Generates a compact structured analysis (~500 output tokens) before the final SQL query, rather than long reasoning chains.
- Training Methodology: Trained on 50,572 execution-verified structured-CoT traces, ensuring high accuracy. It also incorporates column descriptions during training, which significantly boosts performance.
Intended Use & Limitations
- Primary Use Case: Optimized for SQLite text-to-SQL tasks, particularly analytical questions with evidence hints, as evaluated on the BIRD benchmark.
- Efficiency Benefits: Offers cost-effective, on-premises data processing with rapid response times compared to larger models.
- Limitations: Not instruction-tuned for general chat. While highly capable for its niche, its performance on the most challenging reasoning tasks (50.9% EX on BIRD's challenging tier) still reflects the bounds of a 2B parameter model compared to much larger systems.