This project is a Text-to-SQL Query System with Multi-Source Data. It converts natural language queries into SQL queries that can retrieve data from MySQL databases . The system leverages machine learning models (specifically, the T5-small model) to facilitate intuitive query generation.
I explored several text generative models, including Olama, BERT, T5-base, Seq2Seq, and T5-small. Due to my laptop's specifications, I had to choose a model that was compatible without consuming too much time or GPU resources. I ultimately selected T5-small.
However, I encountered an issue where the model generated correct SQL queries but was unable to relate them properly to the database. To address this, I fine-tuned the model on the WikiSQL dataset. Initially, I achieved 92% Rouge score after the first training session. Unfortunately, I could not save the model, necessitating a second round of training. Due to limitations on Google Colab, the training process was interrupted midway, taking around five hours to train for just half the epochs. Each time, the session was disconnected, which forced me to save the partially trained modelLink Text, but it still resulted in errors.
While I have the ML model ready, its performance is limited by the available GPU resources. I then explored other pre-trained models on Hugging Face, but their accuracy matched that of my half-trained model, making them unhelpful. I also discovered the pandas-nql library, which is promising but requires a ChatGPT key. I believe that with access to a proper GPU, I can complete the remaining work to achieve better accuracy.
| Metric | Zero Shot Model | Fine-tuned Model |
|---|---|---|
| Rouge-1 | 0.03197 | 0.92336 |
| Rouge-2 | 0.00500 | 0.88633 |
| Rouge-L | 0.03070 | 0.91765 |
| Rouge-Lsum | 0.03121 | 0.91821 |
For the backend, I utilized Flask to create a REST API, connecting it to my SQL database where my files are stored. I used Postman to test the functionality of my API. My approach effectively searches through the database, and the 92% Rouge score machine learning model is included in this repository under the name fine_tuned_model.ipynb, along with the Flask app in app.py.
The model has been fine-tuned using a P100 GPU notebook updated https://github.com/Darshikartisto/Text_To_SQL/edit/main/texttossql-ipynb.ipynb, and the complete version is now available. It is fully operational and can be tested using FastAPI for integration. Further updates and enhancements will be added soon.
-
Darshika (Darshikartisto)
-
Vaishali (vaishali312003) (Kaggle Notebook)
My biggest challenge was fine-tuning the model without access to a proper GPU. Over the past three days, I kept myself busy tackling this issue.
On the first day, I set up the environment, connected to my SQL database, and researched various models that would fit best with my laptop's specifications and the requirements of the project.
On the second day, I developed the API in app.py and successfully connected it to my database.
By the third day, I focused on fine-tuning my model, which took around five hours for the half-trained model.
Utilizing a better GPU, we successfully completed model training with improved optimization.
📅 Update: 2nd March 2025
| Step | Training Loss | Validation Loss |
|---|---|---|
| 500 | 0.060800 | 0.042019 |
| 1000 | 0.040200 | 0.030493 |
| 1500 | 0.036500 | 0.025905 |
| 2000 | 0.030900 | 0.022786 |
| 2500 | 0.024600 | 0.020304 |
| 3000 | 0.025800 | 0.018751 |
| 3500 | 0.028700 | 0.017546 |
| 4000 | 0.022300 | 0.016787 |
| 4500 | 0.023800 | 0.016067 |
| 5000 | 0.020200 | 0.015377 |
| 5500 | 0.019500 | 0.014977 |
| 6000 | 0.020900 | 0.014771 |
| 6500 | 0.024200 | 0.014634 |
| 7000 | 0.023900 | 0.014581 |
✅ Training completed successfully! 🎉
- CPU Time: User - 2h 24min 29s, System - 27min 16s
- Total Execution Time: 2h 51min 46s
- Wall Time: 2h 51min 37s
- Enhance model accuracy with additional fine-tuning
- Deploy the model for real-time SQL query generation
- Optimize performance for lower latency
To run the Text-to-SQL Query System with Multi-Source Data, follow these steps:
Clone the Repository: Clone the project repository to your local machine then pip install -r requirements.txt , connect to sql database , you may connect yours and then python app.py https://github.com/user-attachments/assets/50261a4a-f954-4cac-827c-a3ebb7b08892