179 questions
No questions match those filters.
← All questions
LLMOps & ProductionMedium· Monitoring· Training Health· Metrics
What training health metrics do you log beyond the loss...
This is one of the questions in the full AI/ML interview bank. Pro unlocks all 1789 questions; Premium includes the same bank plus the highest daily Practice limit.
See plansThe core set of metrics includes loss, pre-clip grad_norm, log_Z, loss_ema_ratio, and nan_frac. Each provides specific insight into training health and should have defined thresholds to trigger alerts or actions. When a run exhibits unexplained instability, add per-layer gradient norm histograms, sampled every 100 steps to limit overhead, to localize the specific layer causing the issue.