Clone Me 🧑

Quickstart 🚀

Clone Me makes it easy to create a personal AI clone by fine-tuning the DeepSeek-R1-Distill-Llama-8B model using the Unsloth library.

Take a peek at the example notebook, or follow these steps to get started:

Step 1: Choose Your Dataset

1. Use Sample Data (Quickstart):
   • Sample data has already been generated for you and is ready to use. The notebook will automatically load this data.
2. Create Your Own Dataset:
   • Head to the Clone Me Website to:
     • Answer pre-built questions.
     • Upload your custom questions in CSV format.
     • Generate training data for your AI clone.

Step 2: Environment Setup

1. Ensure you have access to a GPU environment (Google Colab recommended).
2. Set up your Hugging Face account and API token.
3. (Optional) Create a Weights & Biases account for monitoring training metrics.

Step 3: Fine-Tune the Model

1. Open the Jupyter notebook in ./python/notebooks.
2. Load your dataset (sample or custom).
3. Follow the instructions in the notebook to fine-tune the model.

---

Overview

Personal AI Clone Creation using DeepSeek and Unsloth

This project demonstrates how to fine-tune the DeepSeek-R1-Distill-Llama-8B model using the Unsloth library. The model is optimized for faster, lightweight training and inference, making it a great candidate for experimentation on systems with minimal resources (like Colab). The training uses Complex CoT (Chain of Thought) reasoning and personal data to fine-tune the model for personalized outputs.

Key Features

• Model Distillation: The model is based on distillation from Llama architecture. Basically, take the advanced reasoning from DeepSeek-R1 and integrate it into the efficient Llama model.
• 4-bit Quantization: Efficient resource usage with minimal loss in performance.
• Integration with Weights & Biases: Monitor training metrics in real time.

Requirements

• A GPU environment (Google Colab recommended).
• Hugging Face account with API token.
• Weights & Biases account (optional, for training metrics).
• Personal Q&A dataset in CSV format with columns: Question, Complex_CoT, Response.

---

Artifacts

• Model: DeepSeek-R1-Distill-Llama-8B model fine-tuned on your personal dataset. See my example model on Hugging Face: tdfacer/clone-me-gguf.
• Training Data: Personal Q&A dataset generated using the Clone Me Web App. See my example dataset on Hugging Face: tdfacer/clone_me_generated_sample.

---

Important❗

This project requires training data. To clone yourself:

• Visit the Clone Me Website to create your dataset.
• Alternatively, use the pre-generated sample dataset loaded into this repository or on Hugging Face for convenience.

---

Project Structure

Frontend: Clone Me Web App

• Location: ./frontend/clone-me (or hosted at clone-me-frontend.vercel.app/)
• Description: A Vite/React project for building training datasets via a web interface.
• Features:
  • Custom question upload.
  • Voice-to-text input.
  • Export responses as CSV.

Backend: Python Scripts

• Location: ./python

• Description: Core Python scripts and resources for LLM fine-tuning, dataset generation, and uploading datasets.

  • Notebooks: Jupyter notebooks demonstrating fine-tuning an LLM (e.g., DeepSeek-R1 distilled on Llama 8B).
  • Data: Example question datasets and generated training data.
  • Scripts:
    • clone_me_qa_generator.py: Generates a user persona and answers training questions using an LLM.
    • upload_dataset_to_hub.py: Uploads datasets to the Hugging Face Hub.

---

Development Workflow

1. Frontend

   • Navigate to ./frontend/clone-me.
   • Run npm install and npm start to launch the development server.

2. Python Scripts

   • Set up your Python environment in ./python.
   • Use the provided scripts to generate datasets or upload them to Hugging Face.

3. Fine-Tuning

   • Follow the instructions in the Jupyter notebook to fine-tune the model.

---

Contribution

Contributions are welcome! Feel free to open issues or submit pull requests to improve the project.

---

Tips and Tricks

Push your model to HuggingFace Hub (make sure to choose the gguf version)

#model.push_to_hub_gguf("<your_username>/<your-model-name>", tokenizer, quantization_method = "q4_k_m")
model.push_to_hub_gguf("tdfacer/clone-me-gguf", tokenizer, quantization_method = "q4_k_m")

Download and install ollama Start ollama with the following command

ollama serve

Now run your model with the following command

# ollama run <your_username>/<your-model-name>
ollama run hf.co/tdfacer/clone-me-gguf