Issue #, if available:

Description of changes:

• Skip temperature scaling if zero or negative temperature. A negative temperature could be selected if the model has a worse validation score than a model that predicts a constant 0.5 value. This can happen when there are very few rows of validation data and is an artifact of validation overfitting.
• increase learning rate by 10x to make convergence faster, reduce max iters by 5x to compensate. We can afford a more aggressive learning rate now that we pick the best iteration instead of the final iteration, so divergence is less problematic.

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.