mvishnukarthik/NatraZ ? reverse-engineered prompt

Build me a clean version of this project that can train and test the NatraZ model from the notebooks and the pipeline script.

I want one simple workflow that lets me run the training, save the model, and then test it on new inputs without having to dig through a bunch of notebooks. Keep the existing idea and files in mind, and make it easy for me to open, run, and understand. If you need to check current best practices or docs online, go ahead.

Please organize it so the main steps are clear, the training output is saved properly, and there is an easy way to load the saved model and see results. Also make sure it works smoothly with the existing checkpoint file if that helps.