weiserlab/TinyLLM ? reverse-engineered prompt

Build me a lightweight framework for training, fine tuning, and running small language models on custom data, especially for tiny devices and sensor based projects.

I want it to handle taking in CSV or Hugging Face datasets, turning them into tokenized training data, splitting them into train and validation sets, then training a small GPT style model from those files. After that, I need an easy fine tuning workflow for different datasets like gesture, localization, and breathing type tasks, with saved results, loss plots, and checkpoints.

Please also make it easy to load the trained model for normal text generation in Python, and give me a way to export it into a format that can run with llama.cpp on edge devices. If you need to check current docs for any of the model conversion steps, go ahead and look them up online.