Select desired test sample filters[2]:
Word increased prediction of "falsehood" label  
Word increased prediction of "falsehood" label absence
Filtered Test Samples:
  1. [0] Current model performance over the entire test set. See relevant about page section for additional context ↩
  2. [1] Local accuracy estimated by sorting & bucketing the test set sigmoid outputs, averaging performance in each bucket. See relevant about page section for more context.↩
  3. [2] The explorer samples 100 statements from each of four confidence buckets: the maximum and minimum positive prediction value buckets for each statement type. The selected tweet/nontweet, max/min bucket is then filtered by the specified confusion matrix class. See relevant about page section for more context.↩