Safety and MLOps
Experiment tracking
Experiment tracking records configurations, artifacts, and results for model runs.
How it works
A run becomes reproducible only when code, data version, environment, and parameters are connected. Comparing runs then reveals which change caused a metric shift.
A concrete example
Illustrative: compare 24 runs across 3 learning rates and 2 batch sizes with 4 random seeds.
Do not confuse it with Observability
experiment tracking covers development runs; observability monitors deployed systems.