Skip to content

Troubleshooting

Training Fails After It Worked Once

The most common cause is stale dashboard state. For example, if you start with SVM and then change the estimator to RandomForest without replacing the search space, the backend receives SVM params such as kernel, C, and gamma. RandomForest does not accept those params, so every trial fails.

Fix:

  1. Change estimator.
  2. Let the dashboard reload estimator defaults from GET /api/v1/estimators.
  3. Validate before starting.
  4. Start only if validation passes.

Invalid Metric For Task

Regression supports rmse. Binary classification supports accuracy and roc_auc.

If validation reports a metric error, select a metric that belongs to the chosen task.

Missing Optional Backend

XGBoost and LightGBM are optional. Install only what you need:

uv sync --extra xgboost
uv sync --extra lightgbm

The backend reports missing optional dependencies before a run starts.

CSV Target Column Not Found

Check that target_column exactly matches a CSV header. Column names are case-sensitive.

All Trials Fail

Look at the trial table. Each failed trial includes the estimator error. Common causes:

  • unsupported parameter name;
  • invalid parameter value;
  • incompatible task/metric;
  • target values unsuitable for binary classification;
  • missing optional backend.

SSE Disconnects

The frontend should fall back to polling GET /api/v1/runs/{run_id}. Polling keeps the status, best params, and final result visible even if the event stream disconnects.