Ary-007/Text-to-sql-llama-3.2

Model Details

This model is a fine-tuned version of Llama-3.2-3B-Instruct designed specifically for Text-to-SQL tasks. It was trained to accept a database schema and a natural language question, and output a valid SQL query along with a brief explanation of the logic. It is lightweight (3B parameters), making it suitable for local deployment on consumer GPUs using 4-bit quantization.

Model Description

1. Base Model: unsloth/Llama-3.2-3B-Instruct
2. Fine-tuning Framework: Unsloth (QLoRA)
3. Dataset: gretelai/synthetic_text_to_sql

Uses

The model was trained using the Alpaca prompt format. For best results, structure your input as follows:

How to Get Started with the Model

import torch
from transformers import pipeline

model_id = "Ary-007/Text-to-sql-llama-3.2"

# Load the pipeline
pipe = pipeline(
    "text-generation", 
    model=model_id, 
    device_map="auto",
)

# Define the schema (Context)
schema = """
CREATE TABLE employees (
    id INT,
    name TEXT,
    department TEXT,
    salary INT,
    hire_date DATE
);
"""

# Define the user question
question = "Find the name and salary of employees in the 'Engineering' department who earn more than 80000."

# Format the prompt exactly as trained
prompt = f"""Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.

### Instruction:
Company Database : {schema}

### Input:
SQL Prompt :{question}

### Response:
"""

outputs = pipe(
    prompt, 
    max_new_tokens=200, 
    do_sample=True, 
    temperature=0.1, 
    top_p=0.9
)

print(outputs[0]["generated_text"])

Training Details

The model was fine-tuned using Unsloth on a Tesla T4 GPU (Google Colab).

Hyperparameters

1. Rank (r): 16
2. LoRA Alpha: 16
3. Target Modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
4. Quantization: 4-bit (Normal Float4)
5. Max Sequence Length: 2048
6. Learning Rate: 2e-4
7. Optim: adamw_8bit
8. Max Steps: 60

Dataset Info

The model was trained on the gretelai/synthetic_text_to_sql dataset, utilizing the following fields:

1. sql_context: Used as the database schema context.
2. sql_prompt: The natural language question.
3. sql: The target SQL query.
4. sql_explanation: The explanation of the query logic.

Limitations

1. Training Steps: This model was trained for a limited number of steps (60) as a proof of concept. It may not generalize well to extremely complex or unseen database schemas.
2. Hallucination: Like all LLMs, it may generate syntactically correct but logically incorrect SQL. Always validate the output before running it on a production database.
3. Scope: It is optimized for standard SQL (similar to SQLite/PostgreSQL) as presented in the GretelAI dataset.

License

This model is derived from Llama-3.2 and is subject to the Llama 3.2 Community License.