SQLNova - LoRA Fine-Tuned Deepseek 8B for Text-to-SQL Generation

SQLNova is a lightweight LoRA adapter fine-tuned on top of Unsloth’s Architecture. It is designed to convert natural language instructions into valid SQL queries with minimal compute overhead, making it ideal for integration into data-driven applications or chat interfaces.

The model was trained on over 100,000 natural language-to-SQL pairs spanning diverse domains, including Education, Technical, Healthcare, and more.

Model Dependencies

Python Version: 3.10
libraries: unsloth
pip install unsloth

Model Highlights

Base model: Deepseek R1 8B Distilled Llama
Tokenizer: Compatible with Deepseek R1 8B Distilled Llama
Fine tuned for: Text to SQL Converter
Accuracy: > 85%
Language: English Natural Language Sentences finetuned
Format: safetensors

General Information

Model type: Text Generation
Language: English
License: MIT
Base model: DeepSeek R1 distilled on Llama3 8B

Model Repository

Hugging Face Model Card: https://huggingface.co/mervp/SQLNova

💡 Intended Uses

Applications

Generating SQL queries from natural language prompts
Powering AI assistants for databases
Enhancing SQL query builders or no-code data tools
Automating analytics workflows

Limitations

While SQLNova performs well in many real-world scenarios Since its a Reasoning Model, there are some limitations:

It may produce invalid SQL for rare or malformed inputs in rarest cases.
Assumes a generic SQL dialect, resembling MySQL/PostgreSQL syntax.

Recommendation for Use of Model

Always validate generated SQL before executing in production.
Include schema context in prompts to improve accuracy.
Use with human-in-the-loop review for critical applications.

Thanks for visiting and downloading this model! If this model helped you, please consider leaving a like. Your support helps this model reach more developers and encourages further improvements if any.

How to Use the Model

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="mervp/SQLNova",
    max_seq_length=2048,
    dtype=None,
)

prompt = """ You are an text to SQL query translator.
             Users will ask you questions in English
             and you will generate a SQL query based on their question
             SQL has to be simple, The schema context has been provided to you.


### User Question:
{}

### Sql Context:
{}

### Sql Query:
{}
"""

question = "List the names of customers who have an account balance greater than 6000."
schema = """
CREATE TABLE socially_responsible_lending (
    customer_id INT,
    name VARCHAR(50),
    account_balance DECIMAL(10, 2)
);

INSERT INTO socially_responsible_lending VALUES
    (1, 'james Chad', 5000),
    (2, 'Jane Rajesh', 7000),
    (3, 'Alia Kapoor', 6000),
    (4, 'Fatima Patil', 8000);
"""

inputs = tokenizer(
    [prompt.format(question, schema, "")],
    return_tensors="pt",
    padding=True,
    truncation=True
).to("cuda")

output = model.generate(
    **inputs,
    max_new_tokens=256, 
    temperature=0.2,  
    top_p=0.9,          
    top_k=50,        
    do_sample=True     
)

decoded_output = tokenizer.decode(output[0], skip_special_tokens=True)

if "### Sql Query:" in decoded_output:
    sql_query = decoded_output.split("### Sql Query:")[-1].strip()
else:
    sql_query = decoded_output.strip()

print(sql_query)