I Fine-Tuned an AI ~100x Smaller Than GPT, And It Won

I fine-tuned a 4B parameter model on a single 16GB GPU, then tested it against GPT-5.6 Sol. On a frozen 1,750-document AI-vs-human text detection benchmark, the fine-tuned Qwen model reached 0.915 Macro F1. GPT-5.6 Sol: 0.795 Prompt-only Qwen: 0.541 Fine-tuned Qwen 4B: 0.915 But the interesting part isn't just that the smaller model won. It's why. In this video, I break down when prompting stops being enough, how LoRA and QLoRA make fine-tuning possible on consumer hardware, what actually gets trained, and a few bugs that nearly made the entire experiment useless. #FineTuning #LocalLLM #QLoRA