Many researches are often presented with having to evaluate a new model on multiple datasets or sets of hyperparameters. Research code is famously known for having bad quality due to it’s nature of being thrown away after the publication. Often times no reproducible environment is set up and everything is crammed into a hand-full of scripts, or even worse, jupyter notebooks. These overloaded scripts often try do everything from data manipulation, training and evaluation at once, but more often just ends up as a giant pile of spaghetti nobody wants to deal with.