training-metrics-analyzer

To detect overfitting from TensorBoard logs, you visualize train_loss and val_loss curves to identify where validation loss stops decreasing and diverges from training loss. This skill marks the best checkpoint and subsequent overfitting onset automatically.

The best way to visualize loss curves for optimal checkpoints is plotting train_loss and val_loss with a log-scaled Y-axis. This highlights divergence points, allowing you to select the checkpoint at minimum validation loss before overfitting begins.

Yes, you can compare validation metrics across multiple TensorBoard runs. This skill analyzes logs organized under runs/version_X across experiments, enabling cross-task comparison of training trends and validation losses to evaluate model performance.

To generate visual plots and reports from TensorBoard logs, you run the analyzer script which extracts metrics and renders visualizations using matplotlib and seaborn. Output plots and audit reports are saved automatically under docs/plots and docs/reports.

Yes, this training metrics analyzer works with VLA model training workflows. It is specifically applied to VLA workflows where TensorBoard logs are organized under runs/version_X, extracting relevant metrics to guide checkpoint selection and reporting.