Text2SQL-Flow: A Robust SQL-Aware Data Augmentation Framework for Text-to-SQL

Technical Summary: TEXT2SQL-FLOW

TEXT2SQL-FLOW is a robust, SQL-aware data augmentation framework designed to address the critical shortage of high-quality, diverse, and structurally complex training data for Text-to-SQL tasks. The paper shifts the focus from traditional algorithm-centric approaches to a data-centric paradigm, providing a scalable pipeline to expand limited seed datasets into large-scale, high-fidelity training corpora.

---

1. The Problem: Data Scarcity and Low Diversity

The performance of Large Language Models (LLMs) in Text-to-SQL tasks is heavily bottlenecked by existing datasets (e.g., Spider, BIRD), which suffer from:

• Limited Scale: Manual annotation is expensive and slow.
• Structural Simplicity: Existing datasets often lack complex SQL patterns (e.g., nested subqueries, window functions, or complex joins).
• Narrow Distribution: Datasets often cover limited database schemas and domains, leading to poor generalization in real-world, heterogeneous environments.
• Semantic Gaps: Synthetic data generation often suffers from error propagation or a lack of alignment between natural language (NL) questions and the generated SQL.

2. Methodology: The TEXT2SQL-FLOW Framework

The authors propose an end-to-end pipeline that enhances data across two primary dimensions: Distribution-level diversity and Query-level diversity.

A. Distribution-level Diversity (Schema Expansion):
To prevent models from overfitting to specific database structures, the framework introduces auxiliary databases. It uses a "Database Warehouse" and a selection algorithm based on schema complexity scores (calculated via table count, column count, and foreign key density) to pick diverse databases that cover a wide range of structural complexities.

B. Query-level Diversity (SQL Augmentation):
The framework uses LLMs to transform existing "seed" SQL queries into more complex versions across six dimensions:

1. Data Value Transformations (DVT): Modifying filters, dates, and thresholds.
2. Query Structure Modifications (QSM): Rewriting queries (e.g., replacing JOINs with EXISTS or using CTEs).
3. Business Logic Changes (BLC): Shifting analytical perspectives or domains.
4. Complexity Enhancements (CE): Adding extra joins or CASE expressions.
5. Advanced SQL Features (ASF): Incorporating window functions and set operations.
6. Performance and Optimization (PO): Restructuring queries for efficiency.

C. Quality Control & Pipeline Operators:
To ensure the generated data is usable, the framework integrates several automated filters:

• SQL Executability Filter: Uses EXPLAIN plans to ensure queries are syntactically valid and executable.
• NL Question Generation: Generates questions in 11 different linguistic styles (formal, colloquial, etc.) to ensure linguistic diversity.
• NL–SQL Correspondence Filter: Uses an LLM to verify that the generated question and SQL actually share the same semantic intent.
• CoT (Chain-of-Thought) Generation: Generates step-by-step reasoning traces to help models learn the logical path from question to SQL.

D. Dual-Use Application:

• For Open-Source LLMs: The data is used for Supervised Fine-Tuning (SFT), including the CoT reasoning paths.
• For Closed-Source LLMs: The framework constructs a high-quality knowledge base for Few-shot Retrieval. The authors propose a Masked Alignment Retrieval method that masks schema-dependent tokens (table/column names) to focus on the structural "skeleton" of the query, enabling more accurate example matching.

3. Key Contributions

• TEXT2SQL-FLOW Framework: A systematic, scalable pipeline for SQL-aware data augmentation.
• SQLFLOW Dataset: A high-quality, diverse dataset comprising 75,386 annotated examples (derived from Spider, BIRD, and EHRSQL).
• Masked Alignment Retrieval: A novel retrieval strategy that improves in-context learning for closed-source models by modeling fine-grained structural alignments.

4. Experimental Results

• Fine-tuning Performance: Models trained on SQLFLOW significantly outperformed those trained on existing synthetic datasets (like SynSQL). For example, fine-tuning Qwen2.5-Coder-7B on SQLFLOW yielded a massive +19.1% improvement on the challenging EHRSQL benchmark compared to the baseline.
• Retrieval Performance: The proposed masked alignment retrieval method outperformed standard semantic similarity and AST-based methods, achieving competitive results on BIRD-dev (61.0%) and Spider-dev (83.5%).
• Robustness: Ablation studies confirmed that every component—from auxiliary database selection to CoT generation—contributes significantly to the final model performance.

5. Significance

This work demonstrates that data quality and structural diversity are more impactful than sheer data volume. By providing a method to "upcycle" existing, high-quality seed data into complex, diverse, and verified training sets, the authors provide a blueprint for advancing Text-to-SQL systems toward real-world, enterprise-grade reliability.