---

name: automl-gs description: "Routes automl_gs tabular AutoML searches and the generated runtime artifacts that those searches produce." disable-model-invocation: true metadata: disco-role: operating license: MIT

automl-gs

Use this root skill when the request names automl_gs, asks for an AutoML search on a CSV plus target field, or points at a timestamped generated folder from a previous search.

What this skill covers

• Starting a bounded automl_gs / automl_grid_search(...) run.
• Choosing between the supported search frameworks: TensorFlow or XGBoost.
• Inspecting the generated model.py, pipeline.py, encoders, metadata, and prediction outputs.
• Troubleshooting package import problems, backend mismatches, and generated-folder runtime issues.

Route to a sub-skill

• Use grid-search for a new search from CSV + target.
• Use generated-artifacts for a completed timestamped folder and its train / predict commands.

Read these references when needed

• workflow-overview for the quick package map, supported outputs, and when to choose each route.
• troubleshooting for install/import issues, backend mismatches, and other cross-cutting failures.
• repo-provenance when checking whether this skill matches the current repository revision or before refreshing it.

Install and smoke check

Install the package in the active environment, then add the backend you actually plan to use:

pip install -e .
python -m pip install xgboost

Use TensorFlow only when you need to exercise the legacy TensorFlow-generated path.

For a quick sanity check, run the bundled helper. Pass the backend you actually installed when you want a backend import check:

python scripts/check_install.py --backend xgboost

Fast routing hints

• If the user says “run automl-gs on this CSV”, “search hyperparameters”, or “pick a framework”, route to grid-search.
• If the user says “train the generated model”, “predict from the exported folder”, or mentions model.py / pipeline.py, route to generated-artifacts.
• If import fails, automl_gs -h fails, or a subprocess launches the wrong Python, read troubleshooting first.

Public surface in one line

automl_gs turns a CSV and target field into a generated model folder plus automl_results.csv; the exported folder then owns training, prediction, and encoder loading.