FineTune-Qwen2.5-0.5B

This repository contains scripts and utilities for fine-tuning, testing, and merging LoRA adapters with the Qwen 2.5-0.5B model. The project leverages Hugging Face's transformers library and PEFT (Parameter-Efficient Fine-Tuning) techniques to efficiently fine-tune large language models.

---

Features

• Fine-Tuning with QLoRA: Efficiently fine-tune the Qwen 2.5-0.5B model using LoRA adapters and 4-bit quantization.
• Model Testing: Test the base model, fine-tuned model, and merged model with various configurations.
• Parameter Inspection: Inspect and analyze the parameters of the base model.
• Model Merging: Merge LoRA adapters into the base model for deployment.

---

Installation

1. Clone the repository:

   git clone https://github.com/your-username/FineTune-Qwen2.5-0.5B.git
   cd FineTune-Qwen2.5-0.5B

2. Install the required dependencies:

pip install -r requirements.txt

Usage

1. Fine-Tuning the Model To fine-tune the Qwen 2.5-0.5B model using LoRA adapter

python train.py

This will save the fine-tuned LoRA adapters in the ./qwen2_5_dolly_qlora directory.

2. Merging LoRA Adapters into the Base Model To merge the fine-tuned LoRA adapters into the base model, run:

python merged_QLoRA_to_Base.py

Dependencies

The project requires the following Python libraries:

• transformers
• peft
• torch
• datasets
• matplotlib
• bitsandbytes

For the full list of dependencies, see requirements.txt.

Acknowledgments

• Qwen 2.5-0.5B: A large language model by Alibaba Cloud.
• Hugging Face Transformers: For providing the tools to work with large language models.
• PEFT: For enabling parameter-efficient fine-tuning.