🔧 LLM Tuner: On-Device Fine-Tuner for Niche Domains

LLM Tuner is a lightweight Gradio-based application that allows you to fine-tune open-source language models (like GPT-2) on niche datasets or your own .txt files, directly on-device using LoRA (Low-Rank Adaptation) without using any cloud dependency, ensuring full data privacy and control. It features a simple interface to do fine-tuning and interact with your adapted model through a chat window.

---

🚀 Features

• ✅ Niche Domain Fine-Tuning
  Fine-tune using predefined datasets in domains like Finance, Legal, Medical, and Education.

• 📂 Custom Dataset Upload
  Upload your own .txt file for custom domain fine-tuning.

• 🔁 LoRA Training
  Parameter-efficient fine-tuning using a backend lora_train.py script.

• 💬 Chat Interface
  Chat with the fine-tuned model via an interactive chat window.

• ⬇️ Download LoRA Adapters
  Download LoRA adapter weights after training for offline usage as a .zip.

---

📁 Project Structure

├── app.py # Main Gradio UI and logic
├── lora_train.py # LoRA training script
├── datasets/
│ ├── finance.txt
│ ├── legal.txt
│ ├── medical.txt
│ └── education.txt
├── lora_adapter/ # (Created during training)
└── lora_adapter.zip # (Zip generated after training)

---

⚙️ Requirements

Create a requirements.txt file and install all dependencies:

transformers
gradio
torch
shutil
time
logging

pip install -r requirements.txt

🧪 How to Run

Clone or download the project folder.

Ensure the datasets/ folder includes .txt files for niche domains.

Run the app:

python app.py

Open the link in your browser (usually http://127.0.0.1:7860).

🖥️ Usage Guide

Option 1: Niche Domain Fine-Tuning Click "Select Niche Domain Dataset".

Choose a domain and base model (e.g., GPT-2).

Click on "Train".

Wait a few seconds, fine-tuning will complete.

Download the LoRA adapter for offline usage.

Option 2: Upload Custom Dataset Click "Upload Custom Dataset (.txt)".

Upload your .txt file and choose a base model.

Click "Train".

Fine-tuning will complete in a few seconds.

Download the LoRA adapter.

Chat with Fine-Tuned Model Go to the Chat Interface screen.

Click "Load Fine-Tuned Model".

Start chatting with the fine-tuned model!

📚 Acknowledgements Hugging Face Transformers
Gradio UI Framework
LoRA: Low-Rank Adaptation of Large Language Models

~by Debangan Sarkar, 23117043 :)