LLMRipper

LLMRipper allows you to fine-tune any Hugging Face LLM model without writing any code, using the Transformers architecture.

With LLMRipper.py v1.0, you can perform fine-tuning on either a private or public LLM model and dataset using a single GPU. In the next version, support for multi-GPU training will also be added to the tool’s features. Your training dataset must be in the System-User-Assistant structure.

Features

• You can train public or private Hugging Face models.
• You can provide a CSV or JSON dataset either locally or from Hugging Face.
• You can set the maximum batch size, gradient accumulation steps, and epoch parameters you want for training.
• You can quantize the model and train it using 4-bit or 8-bit QLoRA.
• You can also train in FP16, BF16, or FP32 precision without quantization.
• By using DDP, you can define multiple GPUs for training (Note: This feature will be available in v1.1).
• After training, you can merge the LoRA training outputs into the model and push it to Hugging Face.

Demo

Installation

1. Clone the repository:

   git clone https://github.com/alicankiraz1/LLMRipper.git
   cd LLMRipper
   

2. Install Python Libraries:

   pip install transformers datasets peft bitsandbytes
   
   

3. Run the Code:

   python LLMRipper.py
   

Extras

For training with RTX 5000 Series:

pip unistall torch torchvision torchaudio

pip install --upgrade pip setuptools wheel

pip install --pre torch --index-url https://download.pytorch.org/whl/nightly/cu128

apt-get update && apt-get install -y build-essential cmake