Large language models (LLMs) have demonstrated strong capabilities in translating natural language questions about relational databases into SQL queries. In particular, test-time scaling techniques such as Self-Consistency and Self-Correction can enhance SQL generation accuracy by increasing computational effort during inference. However, these methods have notable limitations: Self-Consistency may select suboptimal outputs despite majority votes, while Self-Correction typically addresses only syntactic errors. To leverage the strengths of both approaches, we propose CSC-SQL, a novel method that integrates Self-Consistency and Self-Correction. CSC-SQL selects the two most frequently occurring outputs from parallel sampling and feeds them into a merge revision model for correction. Additionally, we employ the Group Relative Policy Optimization (GRPO) algorithm to fine-tune both the SQL generation and revision models via reinforcement learning, significantly enhancing output quality. Experimental results confirm the effectiveness and generalizability of CSC-SQL. On the BIRD private test set, our 7B model achieves 71.72% execution accuracy, while the 32B model achieves 73.67%.

Code & Resources

GitHub Repository: https://github.com/CycloneBoy/csc_sql
Hugging Face Collection: https://huggingface.co/collections/cycloneboy/csc-sql-6835c4a52da10c54bbe14f8e
ModelScope Collection: https://modelscope.cn/collections/CSC-SQL-8542177708b643

Framework Overview

Main Results

Performance Comparison of different Text-to-SQL methods on BIRD dev and test dataset:

Models and Datasets on Hugging Face

The following models and datasets related to CSC-SQL are available on Hugging Face:

Model and Dataset	HuggingFace Link
bird train and dev dataset	🤗 HuggingFace
CscSQL-Merge-Qwen2.5-Coder-3B-Instruct	🤗 HuggingFace
CscSQL-Merge-Qwen2.5-Coder-7B-Instruct	🤗 HuggingFace
CscSQL-Grpo-Qwen2.5-Coder-3B-Instruct	🤗 HuggingFace
CscSQL-Grpo-XiYanSQL-QwenCoder-3B-2502	🤗 HuggingFace
CscSQL-Grpo-Qwen2.5-Coder-7B-Instruct	🤗 HuggingFace
CscSQL-Grpo-XiYanSQL-QwenCoder-7B-2502	🤗 HuggingFace

Usage

This model can be loaded using the transformers library. Below is an example of how to use the model for text-to-SQL generation. For more detailed instructions on training and evaluation, please refer to the official GitHub repository.

python

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

# Load the model and tokenizer
model_id = "cycloneboy/CscSQL-Grpo-Qwen2.5-Coder-7B-Instruct" # Example 7B model from the project
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto",
    torch_dtype="auto", # or torch.bfloat16 if supported
    trust_remote_code=True # Required for custom architectures like Qwen2
).eval()

# Prepare your input: natural language question and database schema
question = "What is the average age of students?"
schema_info = """
CREATE TABLE students (
    student_id INT PRIMARY KEY,
    name TEXT,
    age INT,
    major TEXT
);
""" # Replace with actual schema from your database

# Construct the prompt using the Qwen2 chat template format
# The model expects a structured input that includes the schema and question, followed by "SQL:"
formatted_prompt = f"Given the following database schema:
{schema_info}

Generate a SQL query for the following natural language question:
{question}
SQL:"

messages = [
    {"role": "user", "content": formatted_prompt}
]

# Apply the chat template and tokenize
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True # Adds '<|im_start|>assistant
' to prepare for model's response
)

inputs = tokenizer(text, return_tensors="pt").to(model.device)

# Generate the SQL query
generated_ids = model.generate(
    **inputs,
    max_new_tokens=256,
    do_sample=False, # Use greedy decoding for reproducibility
    temperature=0.7,
    top_p=0.9,
    eos_token_id=tokenizer.eos_token_id,
    pad_token_id=tokenizer.pad_token_id,
)

# Decode and print the generated SQL
# Note: The output may contain the original prompt and special tokens. Post-processing might be needed.
output_text = tokenizer.decode(generated_ids[0], skip_special_tokens=True)
print(output_text)

Citation

If you find our work helpful or inspiring, please feel free to cite it:

bibtex

@misc{sheng2025cscsqlcorrectiveselfconsistencytexttosql,
      title={CSC-SQL: Corrective Self-Consistency in Text-to-SQL via Reinforcement Learning}, 
      author={Lei Sheng and Shuai-Shuai Xu},
      year={2025},
      eprint={2505.13271},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2505.13271}, 
}

cycloneboy/CscSQL-Grpo-XiYanSQL-QwenCoder-7B-2502

CSC-SQL: Corrective Self-Consistency in Text-to-SQL via Reinforcement Learning

Abstract