MahderAI: Train, Fine-Tune, and Deploy AI Models

MahderAI is a project dedicated to exploring the end-to-end process of building and deploying Artificial Intelligence models. It serves as a personal workspace for training models from scratch or fine-tuning pre-existing ones (like Google's Gemma), and then making them accessible through user-friendly web interfaces built with React and Django.

The primary goal is to bridge the gap between model development in frameworks like TensorFlow/Keras and practical application via web deployment.

Note: The application is not available for live view as it requires high-end GPUs, which are costly. Instead, you can clone this code and try it yourself or watch the videos I made running both applications:

• Fine-tuning Gemma2 Models: https://youtu.be/PIFOzK_SzVI?feature=shared
• Telegram Hate Content Analyzer: https://youtu.be/VUNmOHDsebY?si=qDUwXu134sNOqSts

Overview

This repository contains the codebase for:

1. Backend API: Built with Django and Django REST Framework to serve AI models, handle training requests, and manage data.
2. Frontend Interface: Developed using React to provide an interactive user experience for interacting with the deployed models and initiating fine-tuning tasks.
3. AI Models: Integration of TensorFlow/Keras models, including custom-trained ones and fine-tuned versions of larger models.

Currently, two main applications are showcased within this project.

Features

• Model Training & Fine-Tuning: Infrastructure and examples for training models or fine-tuning existing ones (e.g., Gemma).
• Web Deployment: Exposes trained models via a Django backend API.
• Interactive Frontend: Interact with models and manage tasks through a React single-page application.
• Modular Design: Separate backend and frontend codebases for clarity and scalability.
• Specific Applications: Includes implementations for:
  • Telegram Hate Speech Classification (Amharic)
  • Gemma Model Fine-Tuning UI

Technology Stack

• Backend: Python, Django
• Frontend: JavaScript, React.js
• Machine Learning: TensorFlow, Keras
• Telegram Integration: Telethon (for the Hate Speech Classifier)

1. Telegram Hate Speech Classifier (Amharic)

• Functionality: Analyzes messages fetched from a public Telegram channel (specified by username) to estimate the proportion of hate speech content. It is specifically trained and optimized for Amharic language text.
• Technology: Uses a custom Transformer-based binary classification model built with TensorFlow/Keras and a SentencePiece tokenizer trained on Amharic data.
• Training Details: The process, dataset information, and code used to train this specific Transformer classifier model can be found in this separate repository: mahdertesf/Transformer-Classifier
• Frontend Component: frontend/src/pages/TransformerClassifier.js
• Backend API: backend/telegramhate/views.py (endpoint: /api/v1/predictHate)

2. Gemma Model Fine-Tuning Interface

• Functionality: Provides a "No Code" web interface to fine-tune Google's Gemma models (2B, 9B, 27B variants) for various NLP tasks using a user-uploaded dataset (.jsonl format).
• Supported Tasks (UI Options): Classification, Extractive Question Answering, Chatbots, Translation.
• Features: Allows configuration of hyperparameters (epochs, batch size, learning rate, validation split), initiates the fine-tuning process on the backend, tracks progress (basic status), and provides a download link for the resulting fine-tuned model checkpoint (.keras file).
• Frontend Component: frontend/src/pages/TrainModels.js
• Backend API: handled by views within backend/app2/ (endpoints for /api/trainmodel, /api/get_training_status, etc.)

Important Notes

• Missing Model Weights: The pre-trained weights for the Telegram Hate Speech Classifier (backend/telegramhate/model_weight/binary_classifier_weights_v21.h5) are NOT included in this GitHub repository due to file size limitations imposed by GitHub.

• Obtaining the Hate Speech Model Weights: The pre-trained and fine-tuned model weights are available on Kaggle at the following location: 🔗 Transformer Classifier (Pre-trained and Fine-tuned)

  Alternatively, you can train the model yourself using the dataset and instructions provided in the dedicated training repository: mahdertesf/Transformer-Classifier

  Once trained, place the generated binary_classifier_weights_v21.h5 (or similarly named) file into the MahderAI/backend/telegramhate/model_weight/ directory.

• Dataset for Pre-training and Fine-tuning: The dataset used for pre-training and fine-tuning the model is available on Kaggle: 🔗 Amharic Transformer Pre-train and Fine-tuning Data

• Telegram API Usage: Using the Hate Speech Classifier requires valid Telegram API credentials. Ensure you have obtained these from Telegram and configured them correctly in your .env file. Use this feature responsibly and be mindful of Telegram's terms of service and rate limits.

Future Work / Roadmap

• Implement more robust training status tracking and logging for the fine-tuning UI.
• Add support for more base models for fine-tuning.
• Deploy the applications to a cloud platform (e.g., Heroku, AWS, Google Cloud).
• Improve error handling and user feedback across both applications.
• Expand language support for the Hate Speech classifier.

Contributing

Contributions, issues, and feature requests are welcome!

Contact

Mahder Tesfaye - GitHub Profile.