Debugging for systems with poor performance

[pfliu-nlp]

This is one nice feedback from @violet-zct
When we discuss AI debugging, what we focus on now is that we have a system with good results and how to continue to improve it. However, researchers often struggle with what to do when the model results are still significantly poor (so far, wandb plays more role regarding this).