unsloth-finetuning 0.2.0

Professional Enterprise Pipeline for Unsloth LLM Finetuning

---

🚀 LLM Finetuning with Unsloth + LoRA + TRL

Fast, lightweight, and scalable finetuning for open-source LLMs

FINE-TUNING AND INFERENCE NOTEBOOKS:

---

📌 Overview

This repo provides a Professional Enterprise Pipeline for finetuning HuggingFace models using:

• ⚡ Unsloth-optimized kernels (2× faster, 70% less VRAM)
• 🏗️ Modular Core Architecture: Unified ModelRunner and ModelFactory for training and inference.
• 🛡️ Hardware Resilience: Internal HardwareManager for pre-flight memory and compatibility checks.
• 🎯 Dynamic Dataset Support: Automatic format detection (Alpaca, ChatML, ShareGPT).
• 🔄 Flexible Deployment: CLI, No-Code GUI, and a clean Python API.

---

🚀 Quick Start: Usage Modes

This repository supports two primary modes of operation.(create a virtualenv: if using uv)

uv run or uv  before_all_commands

1. Requirements

• Python 3.10+ (Required for modern Torch/Transformers compatibility)
• CUDA 12.1+ (For GPU-accelerated training)

2. Installation

Standard Installation:

pip install .

GPU-Accelerated Installation (Local): Installs with full local CUDA stack to avoid system-wide dependency issues.

pip install .[gpu] --extra-index-url https://download.pytorch.org/whl/cu121

or

pip install -r requirements-gpu.txt

2. Mode 1: Headless CLI (Automation)

The CLI is designed for automated training pipelines and reproducible runs.

Basic Usage (via entry point):

unsloth-cli --model_name_or_path "unsloth/mistral-7b-bnb-4bit" --dataset_name "imdb"

Basic Usage (via script):

python src/cli.py --model_name_or_path "unsloth/mistral-7b-bnb-4bit" --dataset_name "imdb"

Using a Config File:

unsloth-cli configs/example.yaml

📖 Read the Full CLI Manual for detailed reference, including dry runs and all argument options.

3. Mode 2: No-Code Studio (Interactive)-[Beta]

Launch the Gradio-based web interface for an interactive fine-tuning experience.

python scripts/app.py

Open your browser at http://localhost:7860

---

☁️ Run on Google Colab

You can easily run this project on Google Colab by cloning the repository.

1. Open a new Colab Notebook.

2. Run the following in a code cell to clone and install:

   !git clone https://github.com/Sriramdayal/Unsloth-LLM-finetuningv1.git
   %cd Unsloth-LLM-finetuningv1
   !pip install -r requirements.txt
   

3. Run Training (CLI Mode):

   !python src/cli.py --model_name_or_path "unsloth/mistral-7b-bnb-4bit" --dataset_name "imdb"
   

---

📘 Python API Guide

For detailed instructions on using this repository as a Python library (including DataProcessor, configuration, and custom training loops), please refer to the dedicated documentation:

👉 Read the Python API Documentation

---

⏱ Training Time (Estimates)

GPU	Small dataset (50k)	Large dataset (500k–800k)
A100	20–40 min	5–7 hours
RTX 4090	1–2 hrs	18–22 hrs
3090 / 4070Ti	1.5–3 hrs	28–34 hrs
Tesla T4	4–6 hrs	55–70 hrs

Training time scales linearly with dataset size and LoRA rank.

---

⭐ Notes & Tips

• Reduce dataset size if needed:

train_ds = train_ds.shuffle(seed=42).select(range(100_000))

• LoRA rank 64 is high quality; use 16–32 for faster training
• Enable flash-attn for memory efficiency on long sequences

---

📜 License

MIT

---

🔗 Credits

• Unsloth
• HuggingFace Transformers
• TRL
• PEFT