nawnie/ReTrain ? reverse-engineered prompt

Build me a Windows first local training workbench for consumer LLMs that feels simple for a normal person to use.

I want to point it at local dataset folders and local model folders, inspect what is there, check if everything is ready, estimate VRAM, and then plan a bounded training run before anything actually starts. It should support dry runs that make receipts and logs without loading model weights, and real runs only when I clearly confirm them. Keep the app standalone, local only, and make it easy to find training receipts, logs, and TensorBoard summaries in one place.

For the actual training side, support the current paths that are already in scope, full SFT, LoRA, and QLoRA for decoder models, plus full fine tuning for text target models and masked language models. If a mode is not ready yet, show it as blocked instead of pretending it works. Make the interface approachable, with clear status, warnings, and simple actions. If you need current library details, look up the latest docs online and wire it up cleanly.