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

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

The model presented in the paper CSC-SQL: Corrective Self-Consistency in Text-to-SQL via Reinforcement Learning.

Abstract: 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%. The code has been open sourced at this https URL.

Code: The code for CSC-SQL is open-sourced at https://github.com/CycloneBoy/csc_sql.

Introduction

CSC-SQL is a novel method that integrates Self-Consistency and Self-Correction for improved Text-to-SQL generation. It addresses limitations of prior methods by selecting optimal outputs and handling both syntactic and semantic errors. The approach employs Group Relative Policy Optimization (GRPO) to fine-tune SQL generation and revision models, leading to significant enhancements in output quality.

Main Results

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

Models

A collection of CSC-SQL models can be found on Hugging Face: CSC-SQL Hugging Face Collection.

Model and 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

Dataset

The BIRD training and development datasets used can be found here: BIRD Train Dataset.

Quickstart

This section provides instructions on how to use the pre-trained CSC-SQL models.

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, GenerationConfig

model_dir = "cycloneboy/CscSQL-Grpo-Qwen2.5-Coder-7B-Instruct" # Or other released models

def load_model_tokenizer(model_path):
    tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
    tokenizer.eos_token = "<|im_end|>"
    tokenizer.pad_token = "<|endoftext|>"
    tokenizer.eos_token_id = tokenizer.convert_tokens_to_ids(tokenizer.eos_token)
    tokenizer.pad_token_id = tokenizer.convert_tokens_to_ids(tokenizer.pad_token)
    tokenizer.padding_side = "left"

    model = AutoModelForCausalLM.from_pretrained(model_path, device_map='auto',torch_dtype=torch.bfloat16, trust_remote_code=True)
    return model, tokenizer

# Example usage for a natural language question (Text-to-SQL)
# Make sure your input string ends with "<|im_start|>assistant
" for generation
text_list = ["""
<|im_start|>user
Your task is to write a SQLite query given a natural language question and a database schema.
You need to generate the SQL query that answers the question correctly.

For example, to find out the names of all the songs, given:
CREATE TABLE songs (
  song_id INTEGER PRIMARY KEY,
  song_name TEXT
);
Question: What are the names of all the songs?
SQL: SELECT song_name FROM songs

To find the artist of the song 'Yesterday', given:
CREATE TABLE songs (
  song_id INTEGER PRIMARY KEY,
  song_name TEXT,
  artist_id INTEGER
);
CREATE TABLE artists (
  artist_id INTEGER PRIMARY KEY,
  artist_name TEXT
);
Question: Who is the artist of the song 'Yesterday'?
SQL: SELECT T2.artist_name FROM songs AS T1 JOIN artists AS T2 ON T1.artist_id = T2.artist_id WHERE T1.song_name = 'Yesterday'

Now, answer the following question.
Question: How many records are there in the table 'songs'?
SQL:
<|im_end|>
<|im_start|>assistant
"""]

model, tokenizer = load_model_tokenizer(model_dir)
inputs = tokenizer(text_list, return_tensors='pt', padding=True, add_special_tokens=False).to('cuda')
input_ids = inputs["input_ids"]
attention_mask = inputs["attention_mask"]
generation_config = GenerationConfig(
        eos_token_id=tokenizer.eos_token_id,
        pad_token_id=tokenizer.pad_token_id,
        temperature=0.1,
        max_new_tokens=512,
        num_return_sequences=1,
        num_beams=1,
        top_p=0.95,
        do_sample=False
)
outputs = model.generate(
        inputs= input_ids,
        attention_mask=attention_mask,
        **generation_config.to_dict()
)
gen_text = tokenizer.batch_decode(outputs[:, input_ids.shape[1]:], skip_special_tokens=True)
print(gen_text[0])

# Expected output: SELECT count(*) FROM songs

Citation

If you find our work useful or helpful for your R&D works, please feel free to cite our paper as below.

@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}, 
}