JPVercosa/machado-transformer ? reverse-engineered prompt

Build me a small project that trains a character level transformer to generate new text in the style of Machado de Assis.

I want it to use a text dataset from Machado works, let me train on either the full collection or a subset like romances or chronicles, and then generate sample text from a saved model. Keep it simple enough to run in Google Colab or locally, with a clear training script and a separate inference script. I’d also like the model settings to be easy to change in one config file, so I can tweak things like training length and model size without digging through the code.

If you need to, look up the current best practices from the PyTorch or Karpathy style transformer examples online. Please make it easy to save and load models, and include a basic way to test generation with a chosen number of tokens.