About hyperparameter-triage
A skill for diagnosing training failures and narrowing hyperparameter tuning to plausible causes. It ranks candidate settings and lays out a minimal search plan with stop conditions.
Key features
- 01Define the objective and failure mode
- 02Rank candidate hyperparameters
- 03Classify knobs by mechanism
- 04Plan a minimal local sweep
- 05Set stop and continue criteria
Use cases
- 01Choosing settings to investigate after a training failure
- 02Reducing an overly broad hyperparameter search
- 03Deciding when to stop a tuning sweep
Install
Copy the command and run it in a terminal.
$
npx skills add https:/After installing, type/hyperparameter-triagein Claude Code to confirm it loaded.
Details
From the project READMEHyperparameter 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
- State the objective and failure mode.
- List high-impact candidate hyperparameters.
- Classify knobs by mechanism: optimization, regularization, capacity, data/sampling, schedule.
- Rank them by expected leverage and interaction risk.
- Propose a minimal search plan.
- 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-auditorortraining-run-forensicsbefore tuning.
Output
Return:
- failure diagnosis;
- ranked hyperparameters with rationale;
- minimal sweep table;
- fixed controls;
- stop/continue criteria.
Comments
Markdown supported