Ultimate Fine-Tuning Guide Covers LoRA, QLoRA

This is the kind of guide people actually want: opinionated, end-to-end, and focused on getting a model from raw weights to a locally usable GGUF file.

• –Strong practical value in covering the boring failure points: CUDA/driver setup, dataset structure, adapter merging, and quantized export
• –The stack is intentionally narrow, which makes it easier to follow but less portable if you are on AMD or a multi-GPU setup
• –Choosing ms-swift plus llama.cpp makes the workflow concrete and reproducible instead of hand-wavy
• –The Qwen3-0.6B example keeps the barrier low, but the writeup is clear that scaling to larger models quickly becomes a VRAM problem
• –The biggest value is probably for developers who know the concepts but want a working recipe they can adapt