Fine-Tuning Llama 3.1 8B for Free on Colab: QLoRA, Unsloth, and a Compression Claim That Doesn't Hold Up

An 8-billion-parameter model usually needs tens of gigabytes of VRAM just to load, let alone retrain. But with 4-bit quantization and LoRA combined, fine-tuning a model that size can run on Colab's fr