What's new in Cometa

Improved drift detection accuracy across all supported frameworks. The monitoring algorithm now accounts for varied data distributions and training patterns, reducing false alerts and delivering more reliable run summaries. Background logging has also been optimized to use less compute overhead throughout training.

Added a redesigned metrics dashboard with cleaner charts, improved color contrast, and a new weekly summary card. Users can now see loss, accuracy, and GPU usage in one glance without having to switch between screens. The layout adapts better to different screen sizes.

Cometa now syncs with PyTorch and TensorFlow automatically. Training logs, checkpoint data, and performance estimates flow between frameworks without any manual input. Users who already train with a custom pipeline will see their full run history pulled in on first sync.

Fixed a bug causing training streaks to reset incorrectly after midnight in certain time zones. Resolved a crash affecting users with large dataset histories on older configurations. Notification delivery timing has also been corrected for users with custom alert schedules.

Cometa 2.0 is here. This release brings a completely rebuilt monitoring engine, a new goal system with daily and weekly targets, run counters, and a redesigned onboarding flow that gets users set up in under two minutes. Performance is significantly faster across all supported configurations.