vandijklab/BrainLM ? reverse-engineered prompt

Build me a BrainLM project that can train a masked prediction model on brain scan data and then use it for simple downstream predictions and reconstructions.

I want the repo to include a clear notebook based workflow for preprocessing the raw fMRI data, getting it into the right shape, and then running inference with a pre trained model. It should support loading the provided weights, doing zero shot style predictions, and showing examples like classifying or regressing from the learned brain features, plus reconstructing a whole recording.

Also make it possible to launch training from a single script, with a setup step that installs the needed Python environment and checks that PyTorch and the Hugging Face tools work. If you need to, look up the current docs online for any tricky model or attention setup details.

Please make it easy to follow from start to finish, with a tutorial notebook, example scripts, and a simple way to plug in processed dataset files.