ihebaker10/text2sql-7b-v3-16

Model Overview

The ihebaker10/text2sql-7b-v3-16 is an 8-billion parameter causal language model, derived from the Arctic-Text2SQL-R1-7B base, specifically fine-tuned for text-to-SQL generation. It processes natural language questions and database schema context to produce PostgreSQL-style SQL queries.

Key Capabilities

• Text-to-SQL Generation: Converts natural language into executable SQL, including support for pgvector retrieval examples using embed_query().
• Structured Output: Generates SQL within a defined <think>...</think><answer>SQL</answer> format.
• Fine-tuned Performance: Utilizes a three-stage PEFT (LoRA) fine-tuning process, including SQL-only warm-up, mixed SQL + pgvector SFT, and GRPO refinement.
• Evaluation Metrics: Achieves 100% format and 99.2% SQL syntax accuracy on a combined validation set, with an 81.8% F1 score for pgvector usage.

Intended Use Cases

• Automated SQL Query Generation: Ideal for applications requiring dynamic SQL query creation from user input.
• PostgreSQL Environments: Specifically trained to generate PostgreSQL-compatible SQL.
• Vector Search Integration: Supports use cases involving pgvector for similarity search queries.

Limitations

• Generated SQL requires validation before execution.
• May undercount semantically equivalent queries due to strict exact-match evaluation.
• Lower pgvector recall suggests some retrieval-style prompts might default to plain SQL.