EuricoGVP/Qwen3.5-4b_text2sql-bird

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 21, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

EuricoGVP/Qwen3.5-4b_text2sql-bird is a 4.56 billion parameter decoder-only language model, fine-tuned from Qwen/Qwen3.5-4B by EuricoGVP. This model specializes in converting natural language questions into SQLite queries, achieving a significant improvement in execution accuracy on the BIRD-SQL benchmark. It is designed for research in text-to-SQL, natural-language interfaces for SQLite databases, and as a component in data-analysis agents.

Loading preview...

Qwen3.5-4b_text2sql-bird: Specialized Text-to-SQL Model

This model, developed by EuricoGVP, is a 4.56 billion parameter decoder-only language model fine-tuned specifically for text-to-SQL tasks. It is based on the Qwen/Qwen3.5-4B architecture, with a merged LoRA adapter, and excels at generating SQLite queries from natural language questions.

Key Capabilities

  • High Accuracy Text-to-SQL: Achieves 53.78% execution accuracy on the BIRD dev set with schema-only prompts and 55.74% when column descriptions are included, representing a gain of up to 9.39 percentage points over the base model.
  • SQLite Specialization: Optimized exclusively for generating queries for SQLite databases.
  • Prompt Format Adherence: Designed to work with a specific system and user prompt structure for optimal performance.

Good For

  • Natural Language Interfaces: Building applications that allow users to query SQLite databases using plain English.
  • Text-to-SQL Research: Serving as a robust baseline or component for further research and development in the text-to-SQL domain.
  • Data Analysis Agents: Integrating into agents that require the ability to translate natural language requests into executable SQLite queries for data retrieval.

Limitations

  • SQLite Only: Not suitable for other SQL dialects (PostgreSQL, MySQL, T-SQL).
  • No General-Purpose Tasks: Not designed for chat, general reasoning, or other broad LLM tasks.
  • Query Validation Recommended: Generated queries should be reviewed or executed in a sandboxed environment, especially for production data, due to potential for non-existent column references or destructive operations.

Finetunes
1 models