Create an experiments.json from `train models`

[hackalog]

Clean up the training loop. concatenate all metadata and save it as experiments.json