Text-to-SQL Query Generation

This project fine-tuned LLaMA-3.2-3B model for generating SQL queries from natural language inputs. The model was fine-tuned using QLoRA (Quantized Low-Rank Adaptation) to efficiently update parameters while minimizing resource usage. It enables users to interact with databases without requiring SQL expertise.

📌 Repository Links

• GitHub: Text-to-SQL_Query-LLM
• Hugging Face Models:
  • QLoRA Model: sai-santhosh/text-2-sql-Llama-3.2-3B
  • GGUF Model: sai-santhosh/text-2-sql-gguf

---

📝 Overview

This project fine-tunes LLaMA-3.2-3B to convert natural language questions into SQL queries, helping users query complex databases with ease. It leverages contextual understanding through CREATE TABLE schemas to improve SQL query accuracy.

🔹 Model Input Format

The model requires two inputs:

• Question: The natural language query, e.g., "List all customers with orders over $500."
• Context: Table schema(s) provided as CREATE TABLE statements to help the model understand database structure.

🔹 Use Cases

• Conversational AI: Enables chatbots to answer database-related queries.
• Educational Tools: Helps users learn and practice SQL with real-world examples.
• Business Intelligence: Simplifies querying large databases for insights.

---

⚙️ Model Details

• Base Model: LLaMA-3.2-3B
• Fine-tuning Method: QLoRA (Quantized Low-Rank Adaptation)
• Task: Text-to-SQL query generation
• Framework: Hugging Face Transformers
• Inference Support: Compatible with LM Studio, Ollama, and GGUF-compatible tools.

---

📦 Installation & Setup

To install the necessary dependencies:

pip install -q -U transformers bitsandbytes accelerate

🚀 Usage

1️⃣ Python API Usage

⚠️ Ensure GPU availability for optimal performance

Clone the repository:

git clone https://github.com/SaiSanthosh1508/Text-to-SQL_Query-LLM
cd Text-to-SQL_Query-LLM

Load the model and generate SQL queries:

from transformers import AutoModelForCausalLM, AutoTokenizer
from text_sql_pipeline import get_sql_query

# Load Model
model = AutoModelForCausalLM.from_pretrained("sai-santhosh/text-2-sql-Llama-3.2-3B", load_in_4bit=True)
tokenizer = AutoTokenizer.from_pretrained("sai-santhosh/text-2-sql-Llama-3.2-3B")

# Example Query
question = "List all employees in the 'Sales' department hired after 2020."
context = "CREATE TABLE employees (id INT, name TEXT, department TEXT, hire_date DATE);"

get_sql_query(model, tokenizer, question, context)

For multiple tables:

question_5 = "Get the names of products that were ordered by customers in New York who spent more than the average amount."
context_5 = "CREATE TABLE Customers (customer_id INTEGER, name VARCHAR, city VARCHAR); \
             CREATE TABLE Orders (order_id INTEGER, customer_id INTEGER, amount INTEGER); \
             CREATE TABLE Products (product_id INTEGER, name VARCHAR); CREATE TABLE Order_Items (order_id INTEGER, product_id INTEGER);"

get_sql_query(model,tokenizer,question_5,context_5)

Output

2️⃣ Command-Line Interface (CLI) Usage

Run the model using the command line:

python generate.py -q "Find the zip code where the mean visibility is lower than 10." \
-c "CREATE TABLE weather (zip_code VARCHAR, mean_visibility_miles INTEGER);"

python generate.py -q "Find all cities with temperatures above 90°F." \
-c "CREATE TABLE weather (zip_code VARCHAR, city VARCHAR, temperature INTEGER);" \
-c "CREATE TABLE population (city VARCHAR, population INTEGER);"

🤗 HuggingFace Spaces Model Inference