Active Learning Improvements

[sonalgoyal]

Knowing when to stop labeling is important to building the right models. To help the user, here are a few ideas

• expose model metrics like accuracy and confusion matrix. These can go as part of the documentation
• see if changing sample size leads to better experience and accuracy
• stratified sampling instead of sampling around 0.5? We had a case where the model went rogue and took a longer cycle to converge