Normalize scores

[tokebe]

BTE result scores currently have no upper bound. These scores should be normalized and bound to [0,1].

@andrewsu proposed a sigmoid function to cause scores to asymptotically approach 1. This makes a lot of sense, given that NGD scoring is unbound. We can simply clamp the low end at 0 as well.

Tagging @andrewsu @newgene for dicsussion.

[andrewsu]

I was thinking of a relatively crude transformation like this (and we can just eyeball the "right" value for the TUNING_PARAM)

https://colab.research.google.com/drive/1FgLR_iOon5W4gJv4-LZtbRkrJDJe7Y4x?usp=sharing

[tokebe]

We could dynamically set the tuning parameter using some statistic of the untransformed set of scores?

[andrewsu]

My gut says we'd want to set a global tuning parameter to allow for the scenario where the answers for one query are truly much better/worse than another. That's how I'd propose we start by implementing it, and then we can examine how it works in practice.

[tokebe]

Deployed to Prod.