Qwen3, Unsloth fuel 5090 style-cloning debate

This is a setup-and-data-quality problem more than a brute-force-parameter problem: on 32GB VRAM, the safest path is a stable QLoRA run on a strong 14B-class base, not a fragile 27B experiment.

• –Unsloth’s docs now explicitly support Windows and Blackwell/RTX 50 GPUs, and the Docker image is the escape hatch if native installs get messy.
• –Their VRAM guidance puts 14B QLoRA well within reach, while 27B is only comfortable in 4-bit training with tighter headroom; full LoRA on 27B is too heavy for this box.
• –For style tuning, the docs recommend `lora_alpha = rank` or `2x rank`, and warn that oversized ranks can overfit; I’d start around `r=16-32` before reaching for `64+`.
• –The voice usually comes more from clean examples, consistent formatting, and decoding choices than from maxing adapter size.