Our CI runners, and Kubernetes are now running jobs successfully, and we're still investigating ways to increase redundancy. We apologize for any disruption this may cause.
Use epsilon for cross entropy of zero-probability features instead of redistributing entropy