04
Nine validation runs cluster near 64–66%; the first run is the outlier
Training and resultsResultThe ten logged accuracies average 64.05% with a 1.76pp sample standard deviation. Because one network instance continues learning across the runs, the sequence is a recorded training benchmark rather than an independent cross-validation estimate.
Recorded validation accuracy by run
Bars are scaled from 55% to 67% to show dispersion.
59.521
63.062
64.603
64.134
63.745
64.826
64.877
65.238
65.809
64.6810
Optimizer
Adam at 0.0004
The training script uses categorical cross-entropy and a padding-aware custom accuracy metric.
Regularization
50% recurrent dropout
The documented architecture applies recurrent dropout to the second and third recurrent layers.
Interpretation
Validation runs are not independent
The code creates the network once, then trains that same instance for one epoch in every recorded run.
| Statistic | Value | Interpretation |
|---|---|---|
| Mean | 64.05% | Matches the recorded 64.0% after rounding |
| Minimum | 59.52% | Run 1 |
| Maximum | 65.80% | Run 9 |
| Sample standard deviation | 1.76pp | Mostly driven by the low first run |
| Validation loss range | 0.5906–0.7077 | Lowest loss occurs in run 9 |