Character LLM Factory 🤖📝🎤

This project demonstrates the complete MLOps lifecycle in practice, from raw data collection to the deployment of a containerized inference service. The goal was to build a "factory" for fine-tuning Large Language Models (LLMs) to learn to speak like a specific character—in this case, Rick Sanchez from Rick and Morty.

🚀 Engineering & MLOps Highlights

This project is not just about fine-tuning; it's a case study on building an end-to-end, robust, and reproducible AI system.

• Automated Data Pipeline (ETL): A scraper built with requests + BeautifulSoup extracts, transforms (cleans and formats), and loads Rick and Morty scripts into a clean text corpus ready for training.
• LLM Fine-Tuning: Leverages the Hugging Face ecosystem (transformers, datasets, accelerate) to fine-tune a base model (distilgpt2) on the custom dataset, teaching it to capture the character's style and personality.
• Optimized Inference Service: The fine-tuned model is served via a high-performance FastAPI service. The model is loaded only once on application startup to ensure low-latency requests.
• Containerization for Portability: The inference application, including the trained model artifact, is packaged into a Docker image, ensuring the service is portable and can run consistently in any environment. Optimized with a .dockerignore file.
• CI/CD & Code Quality: A GitHub Actions pipeline validates the API code's quality (Ruff) and correctness (Pytest) on every push, ensuring service integrity.

🏗️ The MLOps Lifecycle

This project implements the key phases of an MLOps pipeline for custom model creation:

graph TD
    A["Web Source (Transcripts)"] --> B{"1. Data Pipeline (scraper.py)"};
    B -- "Cleans & Formats" --> C["Clean Corpus (rick_corpus.txt)"];
    C --> D{"2. Fine-Tuning Script (trainer.py)"};
    subgraph "Hugging Face"
      D -- Uses --> HF_Base("Base LLM - e.g., distilgpt2");
    end
    D -- Saves --> E["Fine-Tuned Model (rick-llm-final)"];
    E --> F{"3. Inference Service (FastAPI)"};
    subgraph "Docker Container"
        F -- "Loads Model" --> E;
    end
    F --> G["API Endpoint (/generate)"];
    H["User"] --> G;
    G --> H;

Loading

🏁 Getting Started

Prerequisites

• Git, Python 3.9+, Docker Desktop

Option 1: Run the Pre-Trained Service (Recommended)

This option uses the already fine-tuned model included in this repository.

1. Clone the repository:

   git clone [https://github.com/PRYSKAS/character_llm_factory.git](https://github.com/PRYSKAS/character_llm_factory.git)
   cd character_llm_factory

2. Build the Docker image:

   docker build -t rick-llm-service .

3. Run the container:

   docker run -d -p 8001:8001 --name rick-llm rick-llm-service

   Access the API at http://127.0.0.1:8001/docs to interact with the Rick LLM.

Option 2: Recreate the Model (The Full Pipeline)

Follow these steps if you want to run the data pipeline and fine-tuning process from scratch.

1. Clone and install dependencies:

   git clone [https://github.com/PRYSKAS/character_llm_factory.git](https://github.com/PRYSKAS/character_llm_factory.git)
   cd character_llm_factory
   pip install -r requirements.txt

2. Run the Data Pipeline:

   python -m data_pipeline.scraper

   This will generate the rick_corpus.txt file.

3. Run the Fine-Tuning Process:

   python -m finetuning.trainer

   This will train the model and save the artifacts in the rick-llm-final folder.

4. Start the inference service locally:

   uvicorn serving.main:app --reload --port 8001

---