experiments.json…| Experiment | wandb | Checkpoint | Inference · eval sets |
|---|
One row per evaluated checkpoint, grouped by experiment. Each checkpoint has a single
▷ open inference page link; the eval set (subset) — e.g. monitor64, RORD50, PROVE-M/H,
EffectErase VOR, train100 — is chosen from a dropdown inside that viewer (per-clip video grid +
per-example metrics + metric sort). The ckpts chips under each experiment list every saved
checkpoint (not all are evaluated). Use ↔ multi-run compare to stack the same clip across runs
and winner-sort by a metric.