hyper-parm_tuning

What problem does it solve?

Hyperparameter optimization for expensive systems often becomes guesswork that overfits, mixes tuning with confirmation, and produces results that are hard to reproduce or audit; this skill provides a disciplined experimental protocol to prevent those failure modes.

Core Features & Use Cases

• Layerwise tuning with a frozen architecture to avoid searching while the workflow shape is still changing.
• Single-scalar objective discipline so each search loop optimizes one measurable target even when you track dashboards.
• Strict separation of tune vs holdout to ensure the final decision is genuinely confirmed on unseen data.
• Trial persistence and lineage using MLflow as the searchable run ledger (and optional Optuna study for resumability) to make tuning auditable.

Quick Start

Apply hyper-parm_tuning by freezing your system architecture, defining one scalar objective and separate tune/holdout banks, then running layerwise optimization while persisting every trial and finally re-evaluating the best configuration on holdout.