---
license: cc-by-nc-4.0
pipeline_tag: text-generation
library_name: transformers
tags:
  - text-to-sql
  - reinforcement-learning
  - qwen2
  - code-generation
datasets:
  - cycloneboy/bird_train
---

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

This repository contains the models and code for the paper [CSC-SQL: Corrective Self-Consistency in Text-to-SQL via Reinforcement Learning](https://arxiv.org/abs/2505.13271).

📖 [Paper on arXiv](https://arxiv.org/abs/2505.13271) | 💻 [GitHub Repository](https://github.com/CycloneBoy/csc_sql)

## Introduction

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%.

![CSC-SQL Framework](https://github.com/CycloneBoy/csc_sql/raw/main/data/image/csc_sql_framework.png)

## Main Results

Performance Comparison of different Text-to-SQL methods on BIRD dev and test dataset. On the BIRD private test set, our 7B model achieves 71.72% execution accuracy, while the 32B model achieves 73.67%.

<img src="https://github.com/CycloneBoy/csc_sql/raw/main/data/image/csc_sql_result_main.png" height="500" alt="CSC-SQL Main Results">

## Model Checkpoints

The available models are fine-tuned on Qwen2.5-Coder and are accessible on Hugging Face:

| **Model and Dataset**                  | HuggingFace                                                                                 |
| :------------------------------------- | :------------------------------------------------------------------------------------------ |
| CscSQL-Merge-Qwen2.5-Coder-3B-Instruct | [🤗 HuggingFace](https://huggingface.co/cycloneboy/CscSQL-Merge-Qwen2.5-Coder-3B-Instruct)   |
| CscSQL-Merge-Qwen2.5-Coder-7B-Instruct | [🤗 HuggingFace](https://huggingface.co/cycloneboy/CscSQL-Merge-Qwen2.5-Coder-7B-Instruct)   |
| CscSQL-Grpo-Qwen2.5-Coder-3B-Instruct  | [🤗 HuggingFace](https://huggingface.co/cycloneboy/CscSQL-Grpo-Qwen2.5-Coder-3B-Instruct)   |
| CscSQL-Grpo-XiYanSQL-QwenCoder-3B-2502 | [🤗 HuggingFace](https://huggingface.co/cycloneboy/CscSQL-Grpo-XiYanSQL-QwenCoder-3B-2502)  |
| CscSQL-Grpo-Qwen2.5-Coder-7B-Instruct  | [🤗 HuggingFace](https://huggingface.co/cycloneboy/CscSQL-Grpo-Qwen2.5-Coder-7B-Instruct)   |
| CscSQL-Grpo-XiYanSQL-QwenCoder-7B-2502 | [🤗 HuggingFace](https://huggingface.co/cycloneboy/CscSQL-Grpo-XiYanSQL-QwenCoder-7B-2502) |

## Quickstart

You can use this model with the `transformers` library. The `config.json` indicates `Qwen2ForCausalLM` as its architecture.

```python
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, GenerationConfig

model_name = "cycloneboy/CscSQL-Grpo-Qwen2.5-Coder-7B-Instruct"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)
model.eval()

# Example usage for text-to-SQL generation
question = "List the names of all employees."
db_schema = "CREATE TABLE employees (id INT, name TEXT, salary INT);"
prompt = f"Question: {question}
Schema: {db_schema}
SQL: "

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

generation_config = GenerationConfig(
    max_new_tokens=128,
    do_sample=True,
    temperature=0.7,
    top_k=20,
    top_p=0.8,
    repetition_penalty=1.05,
    eos_token_id=tokenizer.eos_token_id,
    pad_token_id=tokenizer.pad_token_id,
)

with torch.no_grad():
    outputs = model.generate(
        **inputs,
        generation_config=generation_config
    )

response = tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)
print(f"Generated SQL: {response}")
```

## Citation

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

```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}, 
}
```