antgroup/Agentar-Scale-SQL-Generation-32B

Important Links

Introduction

We are excited to release the Agentar-Scale-SQL-Generation-32B, the core Reasoning SQL Generator used in our SOTA framework, Agentar-Scale-SQL. Our framework achieved 81.67% execution accuracy on the challenging BIRD benchmark, ranking first on the official leaderboard.

This model is a key component of our "Orchestrated Test-Time Scaling" strategy and has several key features:

• Base Model: It is fine-tuned from Omni-SQL-32B.
• RL-Enhanced Reasoning: The model was further trained using an execution-grounded Reinforcement Learning framework (GRPO) to enhance its intrinsic reasoning capabilities.
• Deep Reasoning: It is engineered to conduct deep, step-by-step reasoning and construct complex, high-accuracy SQL queries.

This model is one of the two main generators in the Agentar-Scale-SQL framework's "Diverse Synthesis" step, working in parallel with an ICL generator to produce a robust pool of SQL candidates.

Model Downloads

Model	Role
Agentar-Scale-SQL-Generation-32B	SQL Generator
Agentar-Scale-SQL-Selection-32B	SQL Selector

Performance

The performance metrics below reflect the entire Agentar-Scale-SQL framework, which uses this Generation model as a key component. The results demonstrate our SOTA performance on the BIRD benchmark.

Methods	EX (Dev)	EX (Test)	R-VES (%)
Agentar-Scale-SQL (Ours)	74.90	81.67	77.00
AskData + GPT-4o	76.14	80.88	76.24
LongData-SQL	74.32	77.53	71.89
CHASE-SQL + Gemini	74.90	76.02	69.94
JoyDataAgent-SQL	74.25	75.74	70.16
TCDataAgent-SQL	74.12	75.74	-
Contextual-SQL	73.50	75.63	70.02
XiYan-SQL	73.34	75.63	71.41

Prompt Template

PROMPT_TEMPLATE = """Task Overview:
You are a data science expert. Below, you are provided with a database schema and a natural language question. Your task is to understand the schema and generate a valid SQL query to answer the question.

Database Engine:
{{ dialect }}

Database Schema:
{{ db_schemas }}
This schema describes the database's structure, including tables, columns, primary keys, foreign keys, and any relevant relationships or constraints.
{% if matched_contents %}
Matched contents:
{{ matched_contents }}
Matched contents presents values related to the question, together with their source table and column, for your reference in SQL generation.
{% endif %}
Question:
{%- if hint %}
{{ hint }}
{{ question }}
{%- else %}
{{ question }}
{%- endif %}

Instructions:
- If Matched contents is provided, you can use it as reference when generating the SQL query.
- Make sure you only output the information that is asked in the question. If the question asks for a specific column, make sure to only include that column in the SELECT clause, nothing more.
- The generated query should return all of the information asked in the question without any missing or extra information.
- Before generating the final SQL query, please think through the steps of how to write the query.

Output Format:
In your answer, please enclose the generated SQL query in a code block:
```sql
-- Your SQL query
```

Take a deep breath and think step by step to find the correct SQL query.
"""

Acknowledgments

If you find our work useful, please cite the Agentar-Scale-SQL paper:

@misc{wang2025agentarscalesqladvancingtexttosqlorchestrated,
      title={Agentar-Scale-SQL: Advancing Text-to-SQL through Orchestrated Test-Time Scaling}, 
      author={Pengfei Wang and Baolin Sun and Xuemei Dong and Yaxun Dai and Hongwei Yuan and Mengdie Chu and Yingqi Gao and Xiang Qi and Peng Zhang and Ying Yan},
      year={2025},
      eprint={2509.24403},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2509.24403}, 
}