YOU NEED TO GET THE WEIGHTS FROM GOOGLE YOURSELF, AND ACCEPT THEIR LICENSE AGREEMENT!

Note that the script provided by DeepMind allows to run the datapipeline in a CPU node by using --norun_inference, and the GPU step with --norun_data_pipeline.

If you do this, the json file that is output of the first step is the input for the second!! Be careful with overwriting and deleting the MSA information.