Fine-Tuning AI for Free — Kaggle + QLoRA Hands-On Guide

I wanted to fine-tune an AI model to sound more human. Not the usual stiff AI tone — something closer to how people actually write on Reddit. Natural, direct, sometimes blunt. So I decided to fine-tune Qwen3-8B on Reddit-style data. The problem was my local PC. Not enough VRAM. So I went looking for a free GPU solution and found Kaggle. Fair warning: I made quite a few mistakes along the way. That's what this post is really about. Why Kaggle Kaggle is known as a data science competition platform, but the key thing is: It gives you free GPU. NVIDIA Tesla T4 (15.6GB VRAM) 30 hours of GPU per week Completely free One thing to know — Kaggle defaults to internet OFF . You can switch it on under Settings → Internet, and turning it on costs nothing extra. I worked with internet OFF, which led to my first mistake. The Full Flow 1. Prepare a public dataset from Hugging Face 2. Connect the model + dataset in Kaggle 3. Run QLoRA fine-tuning 4. Save the adapter and evaluate Step 1 — Data Preparati