Analyze

Where the finished training run lives on disk (an Ultralytics run directory).

The set the run was scored on. Labels drive every deterministic slice.

The run's measured score on THIS eval set.

Distinct from RunMetrics, which is lifted from results.csv and describes the training-val split the run optimised against. Those two are routinely far apart -- an assembled synthetic EO run reads 0.995 on its own val and 0.003 on real IR -- so the training number says nothing about whether the model fails on the supplied eval set. Without this field the harness cannot tell "the model fails here" from "the distribution differs here", and will size a tranche for a model that is already fine.

Only map50 reaches a decision today; the rest are carried for the session record.