Eighteen computer-vision models were trained in one queue on a single power-capped 16 GB Quadro RTX 5000. The queue stopped each run at fixed checkpoints, measured a validation metric, and recorded what the interval cost. This page is that record: every number was written by the runner while the run happened.
Each dot is one training interval. Horizontal position is the wall-clock time it took; vertical position is the change in its validation metric, relative to where the metric already stood. The line at 0.2% is the threshold below which a change is invisible in the output image — on a PSNR of 27 dB it is 0.054 dB.
Four runs spent their final interval going nowhere or backwards. These are not rounding errors — two models ended measurably worse than a checkpoint they had already passed, so the published weights come from earlier in the run.
reached is where the run actually stopped; ceiling is the
budget it was allowed. A run marked early stopped because
its metric flattened, not because it ran out of budget.
Source: the queue runner's log and state file, parsed without modification. Intervals whose metric was unavailable on both sides of the boundary are excluded from the chart and the interval table, which is why the interval count is lower than the number of intervals run.