using-training-optimization

Model learning failures stem from multiple causes—gradient flow issues, learning-rate misconfiguration, or optimizer choice. This skill routes you through diagnostic questions to identify the root symptom, then connects you to the specialist skill for that specific problem, such as gradient-management or learning-rate-scheduling.

Training instability typically traces to learning-rate settings, gradient scaling, or optimizer configuration. Start by clarifying your symptoms—loss spiking, oscillating, or exploding—and this skill routes you to the appropriate optimization specialist to stabilize convergence.

Setting up training optimization involves selecting an optimizer, configuring learning-rate scheduling, managing gradient flow, and tracking experiments. This skill diagnoses your setup requirements and routes you through hyperparameter-tuning, learning-rate-scheduling, gradient-management, and experiment-tracking skills as needed.

Gradient clipping prevents explosion from large gradients; learning-rate adjustment controls convergence speed and stability. This skill asks clarifying questions about your training symptoms—gradient magnitude, loss behavior, convergence rate—then routes to gradient-management or learning-rate-scheduling based on the root cause.

Overfitting shows high train accuracy and low validation accuracy; underfitting shows poor performance on both; optimization issues show stalled or diverging loss. This skill routes you to overfitting-prevention, data-augmentation-strategies, or learning-rate-scheduling based on which symptom you describe.

Hyperparameter selection depends on your optimizer, learning-rate range, batch size, and dataset scale. This skill routes you to hyperparameter-tuning and experiment-tracking skills to systematically search and log configurations, then validates which choices resolve your training symptoms.