CoolFace
Modelpublic

tdfacer/clone-me-gguf

sourceHugging Faceupdated 2y agoView on Hugging Face
1likes50downloads
Model Card

Clone Me 🧑

Fine-Tuned DeepSeek-R1-Distill-Llama-8B

Overview

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. This model is optimized for faster, lightweight training and inference, making it a great candidate for personalized experimentation and deployment.

The training process leverages:

  • —Model Distillation: Fine-tuned from DeepSeek-R1 using Llama architecture for efficient reasoning.
  • —4-bit Quantization: Ensures resource efficiency with minimal loss in performance.
  • —Complex Chain-of-Thought Reasoning: Generates responses with detailed, logical reasoning paths.

Key Features

  • —Fine-Tuned Dataset: Includes personal Q&A datasets for creating personalized clones.
  • —Frontend Integration: Build and refine datasets using the Clone Me Website.
  • —Ready for Deployment: GGUF-compatible format for use with ollama or local inference tools.

Links

How to Use

Quickstart 🚀

Running the Model Locally
  1. 1.Install Ollama
  2. 2.See ollama docs here
  1. 1.Run the Model
  2. 2.Use ollama to serve the GGUF version:
bash
     ollama serve
     ollama run hf.co/tdfacer/clone-me-gguf
Training a Clone
  1. 1.Use Pre-Generated Data
  2. 2.Sample datasets: tdfacer/clone_me_generated_sample.
  3. 3.OR Create Your Own Dataset
  4. 4.Visit the Clone Me Website to generate training data by answering pre-built or custom questions.
  1. 1.Train Using the Colab Notebook
  2. 2.Start training immediately in this Colab Notebook.

Uses

Direct Use

  • —Use as a foundation for creating AI clones with logical, personalized outputs.

Downstream Use

  • —Fine-tune further for specific domains or applications.
  • —Deploy using lightweight inference frameworks like Ollama or llama.cpp.

Out-of-Scope Use

  • —Avoid deploying in high-stakes applications without rigorous testing.
  • —Not intended for applications requiring real-time critical decisions.

Training Details

Training Data

The model is fine-tuned on:

  • —Personal Q&A Dataset: User-generated training data from Clone Me.

Training Procedure

  • —Library: Unsloth for efficient training.
  • —Quantization: 4-bit for low-resource environments.
  • —Hardware: Fine-tuned on GPU environments (e.g., Colab).

Limitations

  • —Bias: Responses may inherit biases from the training dataset.
  • —Accuracy: May struggle with highly technical or ambiguous queries.
  • —Resource Constraints: Fine-tuning requires GPU; CPU-only environments may see slower inference speeds.

Contact

Citation

If you use this model, please cite:

plaintext
@model{tdfacer_clone_me,
  author = {Trevor Facer},
  title = {Clone Me: Fine-Tuned DeepSeek-R1-Distill-Llama-8B},
  year = {2025},
  url = {https://huggingface.co/tdfacer/clone-me-gguf},
}