running-hyperparameter-sweep

What problem does it solve?

This Skill turns ad hoc model tuning into a disciplined hyperparameter sweep, helping you improve performance without accidentally optimizing on the test set or trusting a single noisy run.

Core Features & Use Cases

• Search-space design: Chooses appropriate distributions for learning rate, weight decay, batch size, optimizer, and related knobs.
• Sampler and pruner selection: Recommends practical Optuna or Ray Tune strategies such as TPE, random, ASHA, median pruning, or Hyperband based on compute and dimensionality.
• Compute budgeting: Splits time between the sweep and the final retrain so the best candidate is validated with multiple fresh seeds before winner selection.
• Safety guardrails: Enforces a training-versus-validation-versus-test firewall, flags boundary solutions, and warns when the sweep is too small or the landscape is flat.
• Use case: Ideal for expensive deep learning runs where defaults are weak, manual tuning has stalled, or you need a repeatable process for comparing top configurations.

Quick Start

Use this skill to plan a safe hyperparameter sweep for my model and tell me the search space, sampler, pruner, budget split, and retraining protocol.