hyperparameter-triage

Hyperparameter Triage

Tune the few knobs that are actually plausible culprits.

Intake

Establish:

• target metric and failure mode;
• current config and compute budget;
• training curves and validation behavior;
• known-good baseline settings;
• which knobs have already been tried.

Workflow

1. State the objective and failure mode.
2. List high-impact candidate hyperparameters.
3. Classify knobs by mechanism: optimization, regularization, capacity, data/sampling, schedule.
4. Rank them by expected leverage and interaction risk.
5. Propose a minimal search plan.
6. Define stop conditions for continuing or abandoning the sweep.

Rules

• Prefer informed local sweeps over blind grid search.
• Document what was held fixed.
• Change one mechanism at a time unless testing an interaction deliberately.
• If curves imply a data or implementation bug, route to data-pipeline-auditor or training-run-forensics before tuning.

Output

Return:

• failure diagnosis;
• ranked hyperparameters with rationale;
• minimal sweep table;
• fixed controls;
• stop/continue criteria.