Skip to content

README.md

Latest commit

 

History

History
159 lines (120 loc) · 5.34 KB

README.md

File metadata and controls

159 lines (120 loc) · 5.34 KB
Raw

Simpsons Dialogue Generator

An AI-powered web application that generates something close to authentic-sounding dialogue from The Simpsons using a fine-tuned GPT-2 language model.

Project Overview

This project demonstrates the complete machine learning workflow from data preparation to deployment, showcasing how to fine-tune a language model on a custom dataset using consumer-grade hardware. The model generates character-specific dialogue that captures the personality and speaking style of iconic Simpsons characters.

Model: Hugging Face Space

Features

Interactive Web Interface

  • Character Selection: Choose from various Simpsons characters
  • Location Context: Set the scene with different Springfield locations
  • Prompt Words: Guide the dialogue generation with starter words
  • Advanced Parameters: Fine-tune generation with adjustable settings
    • Temperature (creativity level)
    • Top-k and Top-p sampling
    • Maximum output length
    • Number of outputs
    • Repetition penalty

Technical Stack

Machine Learning

  • Model: GPT-2 (124M parameters)
  • Framework: PyTorch + Hugging Face Transformers
  • Training Hardware: AMD Radeon RX 6950 XT GPU (16GB VRAM)
    • CPU: Intel i5-13400F
    • RAM: 32GB DDR4
    • OS: Linux Mint(22.3 Zena)

Frontend

  • Framework: React 18 with TypeScript
  • Styling: Tailwind CSS
  • Icons: Lucide React
  • API Client: Gradio Client

Deployment

  • Model Hosting: Hugging Face Spaces (Gradio)
  • Frontend: Cloudflare
  • Architecture: Serverless API calls to Hugging Face

How It Works

Training Process

  1. Dataset Preparation: Collected and structured thousands of Simpsons dialogue samples from Kaggle
  2. Data Format: Custom conditioning format: [CHAR:Name][LOC:Location][PROMPT:Word]
  3. Fine-tuning: Trained GPT-2 on local hardware with character/location awareness
  4. Evaluation: Tested for personality consistency and contextual relevance

Inference Pipeline

  1. User selects character, location, and prompt word
  2. Frontend formats input with special tokens
  3. Request sent to Hugging Face Gradio API
  4. Model generates dialogue using configured sampling parameters
  5. Response displayed with formatted output

Special Features

  • Context Conditioning: Model learns to associate characters with their speech patterns
  • Location Awareness: Setting influences dialogue topics and references
  • Controlled Generation: Multiple parameters allow fine-grained output control

Learning Outcomes

This project was designed as a comprehensive learning exercise in:

Hardware & Infrastructure

  • Configuring AMD ROCm for deep learning on RDNA2 architecture
  • Managing GPU memory constraints during training
  • Optimizing training pipelines for consumer hardware

Machine Learning Engineering

  • Fine-tuning pretrained language models
  • Custom tokenization and data preprocessing
  • Implementing conditional text generation
  • Hyperparameter tuning and experimentation

Full-Stack Development

  • Building responsive React applications
  • Integrating ML models via REST APIs
  • Managing asynchronous operations and loading states
  • Deploying serverless ML applications

Realistic Expectations

  • Understanding the resource gap between hobby projects and production AI
  • Learning to work within hardware limitations
  • Balancing model quality with practical constraints

Getting Started

Prerequisites

node >= 18.0.0
npm >= 9.0.0

Installation

# Clone the repository
git clone https://github.com/jonorl/simpsons-GPT.git
cd simpsons-GPT

# Install dependencies
npm install

# Run development server
npm run dev

Environment Setup

The application connects to the Hugging Face Spaces API automatically. No API keys required for basic usage.

Model Parameters

Parameter Default Range Purpose
Temperature 0.8 0.1-2.0 Controls randomness
Top-k 50 1-100 Limits vocabulary selection
Top-p 0.9 0.1-1.0 Nucleus sampling threshold
Max Length 100 20-200 Maximum tokens generated
Repetition Penalty 1.2 1.0-2.0 Reduces repeated phrases

Known Limitations

Model Quality

  • Trained on consumer hardware, not production-scale infrastructure
  • Output may be abstract or nonsensical compared to industry models
  • Limited by ~124M parameters vs billions in modern LLMs

Performance

  • Free-tier Hugging Face Spaces has cold start delays
  • First request may take 30-60 seconds
  • Subsequent requests are faster (server warm)

Content

  • Model trained on TV dialogue, not real-world conversations
  • May occasionally generate unexpected combinations
  • Best used for demonstration and learning purposes

License

This project is for educational and portfolio purposes. The Simpsons characters and content are property of 20th Century Fox/Disney.

Author

Jonathan Orlowski

Acknowledgments

  • Prashant Banerjee (prashant111) for the Kaggle Simpsons dataset
  • Hugging Face for model hosting and Transformers library
  • PyTorch team for the deep learning framework
  • Gradio for the deployment interface
  • AMD for ROCm GPU support
  • The Simpsons for decades of memorable dialogue