makazhanalpamys/soup ? reverse-engineered prompt

Build me a simple Python tool that lets someone fine tune and post train an LLM from one YAML file and one command.

I want the flow to feel easy for a normal person, like install it, pick a template, edit a single config, then run training locally without SSH or a bunch of setup pain. It should work on a laptop GPU with low VRAM, and it should automatically handle the annoying stuff like choosing sane batch sizes, using quantization, and figuring out what the machine can support.

Also include a basic web UI or dashboard so I can watch training progress, see the loss and other useful metrics, and check what config is being used. If there are a few common example recipes for chat style fine tuning, add those too. Make it friendly, clear, and hard to mess up, and look up current docs online if you need to.