Fine-Tuning_Qwen3-4B_with_Unsloth

Fine-Tuning Qwen3-4B with Unsloth Kaggle Notebook

Status: 🚀 Completed
Tech Stack: Python, PyTorch, Unsloth, Hugging Face (Transformers, TRL, Datasets), QLoRA, Groq API, Weights & Biases, ipywidgets.

Overview

ProTutor is a custom 4-billion parameter AI tutor fine-tuned to teach software engineering concepts. This project demonstrates an end-to-end Large Language Model (LLM) distillation pipeline, leveraging a 120B parameter teacher model to generate synthetic training data, which is then used to fine-tune a smaller student model (Qwen3-4B).

Built entirely on a free Kaggle T4 GPU, this pipeline utilizes QLoRA (Quantized Low-Rank Adaptation) and the Unsloth framework to optimize memory usage and accelerate training speed, completing the entire fine-tuning process in under 30 minutes.

Setup & Installation

This project is designed to run natively on a Kaggle Notebook with a T4 GPU (×2) or any local Jupyter environment with at least 16GB VRAM.

1. Environment Setup (Kaggle): Ensure Internet and GPU T4 are enabled in notebook settings.
2. API Keys: Add your free Groq API key to Kaggle Secrets as GROQ_API_KEY.
3. Install Dependencies: Run the following command to install the required stack (Note: Unsloth requires specific PyTorch/CUDA wheels):

   pip install -r requirements.txt
   

Architecture & Flow

1. Synthetic Data Generation (Distillation): - Utilized the openai/gpt-oss-120b model via the Groq API to asynchronously generate ~300 high-quality question-answer pairs.
   • Enforced a strict zero-shot persona (standard English, universal analogies, specific summary formatting) without relying on system prompts at inference.
2. Formatting & Tokenization: - Serialized raw JSON data into the exact ChatML template (<|im_start|>, <|im_end|>) expected by the Qwen instruct model.
3. Model Initialization & QLoRA: - Loaded the Qwen3-4B-Instruct base model in 4-bit precision using Unsloth.
   • Attached trainable LoRA matrices (Rank = 32) to the attention and MLP projections, targeting roughly 1.6% (66M) of the total model parameters.
4. Training Loop: - Executed supervised fine-tuning using TRL's SFTTrainer.
   • Applied train_on_responses_only to mask the loss on user prompts, forcing the model to strictly optimize for the assistant's persona.
   • Tracked training loss and metrics via Weights & Biases.
5. Interactive Evaluation & Export: - Built a custom, interactive chat UI within the notebook using ipywidgets to test the held-out validation set.
   • Exported the final trained weights to GGUF format for local/edge deployment.