Samson Tan

@KhalidAlt @stephenbach Any chance we can get this merged soon? We need the filter to be removed before we can run the evaluations for the paper. Thanks!

@anoff me neither. In fact, I'd rather have more commits if it means the PR is more organized (one commit per logical change). The squashing I refer to is upon...

@anoff if this is something running locally (which it seems to be), I think we can safely let the individual devs decide if they want to use it. We could...

@anoff Yup! We'll have a convention, but whether or not the contributors chooses to use the tool to help them is up to them.

Great points raised here. First thing that came to my mind was feedforward (ideation) and backpropagation (defining issues) :P

> What if we hide the FeedbackResponse object inside the FeedbackResponseAttributes? That way, we can avoid step 4. Would replacing the current *Attribute classes with some sort of wrapper class...

`T5TokenzierFast` does not have byte-fallback + why artificially handicap the slow tokenizer if it could be more efficient (using sets instead of lists and computing the attribute only when it's...