The best model performance is given with its System and User prompts. The model is intended to be used with three inputs: question, evidence, and the database schema.

Required transformers > 4.51.0 to have Qwen3. Make sure to update your transformers installation via pip install --upgrade transformers.

python

import transformers
import torch
model_id = "anonymous-2321/Think2SQL-14B"
pipeline = transformers.pipeline(
    "text-generation",
    model=model_id,
    model_kwargs={"torch_dtype": torch.bfloat16},
    device_map="auto",
)

system_message ="""
You are a data science expert that provides well-reasoned and detailed responses. Your task is to understand the schema and generate a valid SQL query to answer the question.
You first think about the reasoning process as an internal monologue and then provide the user with the answer.
Respond in the following format:
<reasoning>
    ...
</reasoning>
<answer>
    ...
</answer>
""".strip()

user_message = """
Answer the following question with the SQL code. Use the piece of evidence and base your answer on the database schema.
Given the question, the evidence and the database schema, return in the <answer> tags only the SQL script that addresses the question.

Database Engine:
SQLite

Question:
Return the product name, sorted alphabetically and by price in descending order.


Evidence:


Database Schema:
CREATE TABLE products (
    id INTEGER PRIMARY KEY,
    name TEXT NOT NULL,
    price REAL NOT NULL
);

CREATE TABLE customers (
    id INTEGER PRIMARY KEY,
    name TEXT NOT NULL,
    email TEXT NOT NULL
);
"""


messages = [
    {"role": "system", "content": system_message},
    {"role": "user", "content": user_message},
]

outputs = pipeline(
    messages,
    max_new_tokens=4096,
    temperature=0.6,
    top_p=0.95,
    top_k=20
)
print(outputs[0]["generated_text"][-1])

📖 Overview

Think2SQL is a systematic study on injecting reasoning capabilities into Text-to-SQL through Reinforcement Learning with Verifiable Rewards (RLVR). We uncover the critical interplay between reward density, advantage scaling, and model capacity, proposing novel execution-guided dense rewards and optimal scaling strategies. Our 4B-parameter model achieves reasoning capabilities competitive with state-of-the-art models, while providing a comprehensive analysis for optimizing Text-to-SQL reasoning under computational constraints.

Key Contributions:

Execution-guided dense reward function that outperforms binary signals
Analysis of advantage scaling mechanics for models of different sizes
Evaluation of cold start effects and supervised fine-tuning impact
Pareto frontier mapping for training efficiency optimization