Agentar-Scale-SQL: Advancing Text-to-SQL through Orchestrated Test-Time Scaling

📝 Introduction

Agentar-Scale-SQL is a novel framework that leverages scalable computation to significantly improve Text-to-SQL performance on challenging benchmarks. By implementing an Orchestrated Test-Time Scaling strategy, our framework synergistically combines three distinct perspectives to bridge the gap between state-of-the-art models and human expert performance.

<figure>
<img src="./resources/framework.png"
alt="framework"
width="1200"
style="display: block; margin: 0 auto;">
<figcaption style="text-align: center; font-style: italic; color: #555;">
Figure 1: The proposed Agentar-Scale-SQL framework.
</figcaption>
</figure>

⚡️ Performance

🎉 News

• 📝 2026.05.07: Our new paper Every Step Counts: Step-Level Credit Assignment for Tool-Integrated Text-to-SQL is available on arXiv, introducing FineStep for fine-grained step-level credit assignment in tool-integrated Text-to-SQL.
• 🚀 2025.11.27: We are excited to release Agentar-Scale-SQL-Generation-32B on Hugging Face and ModelScope! Simultaneously, we have open-sourced the code for the Light Schema Engine and the Offline Data Preprocessing Pipeline!
• 🎁 2025.09.30: Our paper is available on arXiv.
• 🏆 2025.09.25: We are proud to announce that we have achieved #1 Rank on the official BIRD leaderboard with 81.67% execution accuracy!

🗺️ Release Roadmap

We are committed to continuously improving Agentar-Scale-SQL. Here is our plan for upcoming features and releases.

• Paper
  • [x] Publish the Paper on arXiv.
• Model Releases
  • [x] Release Agentar-Scale-SQL-Generation-32B on Hugging Face and ModelScope.
  • [ ] Release Agentar-Scale-SQL-Selection-32B on Hugging Face and ModelScope.
• Code Releases
  • [x] Release the code for the Light Schema Engine.
  • [x] Release the code for the Offline Data Preprocessing Pipeline.
  • [ ] Release the code for Task Understanding and Generating SQL Candidates with ICL Generators.
  • [ ] Release the code for Generating SQL Candidates with the Reasoning Generator.
  • [ ] Release the code for the Iterative Refinement module.
  • [ ] Release the code for the SQL Selection module.

📂 Directory Structure

Agentar-Scale-SQL/
├── ScaleSQL/                     # Core source code directory
│   └── workflows/                # Main workflow scripts
│       └── config/               # Configuration files
├── ddl_schema.sh
├── requirements.txt              # Dependency list
├── .env                          # Environment variable
├── .env.example                  # Environment variable template
├── .gitignore
├── README.md                     # Current document
├── nltk_data.zip                 # For ddl schema generation

📚 Usage

1. Installation and Environment Settings

1.1 Create Virtual Environment and Install Python Dependencies

conda create -n ScaleSQL python=3.10
conda activate ScaleSQL

1.2 Install PyTorch and Core Dependencies

# Install PyTorch (CUDA 12.1)
pip install torch==2.5.1 torchvision==0.20.1 torchaudio==2.5.1 --index-url https://download.pytorch.org/whl/cu121

1.3 Install Project Dependencies

pip install -r requirements.txt

1.4 Install vLLM (for Inference Acceleration)

pip install https://github.com/vllm-project/vllm/releases/download/v0.8.5.post1/vllm-0.8.5.post1+cu121-cp38-abi3-manylinux1_x86_64.whl

1.5 Download Embedding Model

modelscope download --model sentence-transformers/all-MiniLM-L6-v2 --local_dir ./ScaleSQL/model/all-MiniLM-L6-v2

2. Data Preparation

2.1 Configure Paths

Modify the configuration file: .ScaleSQL/workflows/config/pipeline_config.yaml.
Note that, we need column meaning file in the evaluation. You can find the file in TA-SQL.

dataset_folder: /temp/bird_test  # Change to the actual folder
column_meaning_path: /your_path/column_meaning.json # Change to the actual path

3. Preprocessing Pipeline

3.1 Generate Light Schema

python -m ScaleSQL.workflows.schema_generation --evaluation_type test

Output example: .ScaleSQL/dataset/bird_test_light_schema.json

3.2 Process Training Set Examples and Write to Vector Database

ANONYMIZED_TELEMETRY=False python -m ScaleSQL.workflows.train_skeleton_process

Output path: /tmp/ScaleSQL/chroma/bird_train_skeleton

3.3 Process Database Cell Values and Write to Vector Database

ANONYMIZED_TELEMETRY=False python -m ScaleSQL.workflows.database_cell_process --evaluation_type test

Output path: /tmp/ScaleSQL/chroma/bird_test

3.4 Build BM25 Index (Content-Based) and Generate DDL Schema (Requires Java Environment)

bash ddl_schema.sh

Output example: .ScaleSQL/dataset/bird_test_ddl_schema.json

📦 Try Our Product

Unlock the power of your business data with natural language. We are excited to introduce Data Agent, our cutting-edge ChatBI product designed to transform complex data into clear, conversational insights.

Simply ask questions in plain English, and let Data Agent handle the complex queries for you. No code, no steep learning curve—just instant answers.

• Official Website: For a detailed overview, features, and use cases, please visit our product page:
  https://antdigital.com/products/DataAgent

• Early Access & Feedback: If you are interested in trying our product and providing valuable feedback, please feel free to contact us.

<figure>
<img src="./resources/dingding.png"
alt="dingding"
width="250"
style="display: block; margin: 0 auto;">
<figcaption style="text-align: center; font-style: italic; color: #555;">
Figure 2: The contact information.
</figcaption>
</figure>

📎 Citation

@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}, 
}