bsana1/arrodeio-llm ? reverse-engineered prompt

Build me a Python project that teaches LLMs using Brazilian folk sayings and public domain cordel poetry.

I want two paths in the same repo. One should train a tiny GPT from scratch on a small merged text corpus, with a tokenizer, training script, and a simple generator so I can see it go from nonsense to memorizing the training data. The other should take a pretrained Portuguese GPT 2 model and fine tune it on the same corpus so it can generate readable cordel style text.

Please include scripts to fetch the cordel sources, prepare the dataset, train the tokenizer, train both models, and generate sample text from each. Keep it able to run on a normal laptop CPU, and make the outputs easy to compare so the difference between from scratch training and fine tuning is obvious. If you need to look up current library docs online, go ahead.