illidanlab/HyFL ? reverse-engineered prompt

I want a working Python project that implements a privacy preserving hybrid federated learning setup for financial crime detection, based on the HyFL idea in the README and paper. It should be something I can run locally with Docker, train on the provided financial crime datasets, and then test to generate predictions in the expected CSV format.

Please wire up the full flow so the repo can handle both centralized and federated runs, using the data folder layout described in the README with scenarios, client partitions, train and test splits, and prediction templates. I want sensible training and inference scripts, config where needed, and clear commands so I can build the image, run training, run testing, and package the solution for submission. If anything is missing, add the minimum code and docs needed to make it usable.

Please also make the setup reasonably robust and easy to follow for someone who is not deep into federated learning. You can look up the current docs for the challenge runtime and related framework if needed.