tensorflow/tensor2tensor ? reverse-engineered prompt

Build me a Python library like Tensor2Tensor that makes it easy to train and evaluate deep learning models on different datasets from the command line.

I want a clean package with a trainer, a few example models, a few example tasks and datasets, and simple configs so someone can run one command and start training on something like MNIST or a text task. It should feel like a research library, easy to add new problems, models, and hyperparameter sets, with a small demo or notebook that shows the whole flow from install to training to evaluation.

Please make it usable out of the box, with a README that explains the basics, sample commands, and how to plug in custom data. If you need to check current docs for any TensorFlow details, feel free to look them up online.