Cli And Model Registry

---
name: cli-and-model-registry
description: "Guides Autodistill CLI runs, model registry inspection, plugin
  selection, and safe dry-run troubleshooting."
disable-model-invocation: true
metadata:
  disco-role: operating
license: Apache 2.0
---

# Autodistill CLI and Model Registry

Use this sub-skill when a task asks for an `autodistill ...` command, model alias selection, plugin package requirements, model support matrix interpretation, registry inspection, or a safe dry run before downloads/training.

The core CLI orchestrates plugin base models and target models. It can import or install plugin packages, label an image folder, train a target model, and optionally upload to Roboflow. Treat full CLI runs as side-effectful unless the user explicitly approves plugin installs, model downloads, training time, hardware use, output paths, and credentials.

## Quick Route

- **Build or explain CLI commands:** read [CLI reference](references/cli-reference.md) for verified options, boolean value quirks, JSON ontology quoting, and safe examples.
- **Choose base/target aliases:** read [model registry](references/model-registry.md) for supported aliases, package naming, default target checkpoints, and model matrix boundaries.
- **Debug CLI failures:** read [troubleshooting](references/troubleshooting.md) for invalid ontology JSON, unsupported aliases, missing plugins, `--model_type` mismatch, Roboflow credentials, and training side effects.
- **Check CLI help safely:** run [scripts/autodistill_cli_smoke.py](scripts/autodistill_cli_smoke.py). It runs help and optional ontology JSON validation only.
- **Inspect registry safely:** run [scripts/inspect_model_registry.py](scripts/inspect_model_registry.py). It lists aliases and installed/missing plugin modules without installing anything.

## Minimal Safe Checks

```bash
python scripts/autodistill_cli_smoke.py \
  --ontology-json '{"milk bottle": "bottle"}'

python scripts/inspect_model_registry.py
```

If these pass, the core CLI/registry is importable. They do **not** prove that a concrete plugin can download weights, run inference, train, or upload.

## Full CLI Pattern

```bash
autodistill images \
  --base grounding_dino \
  --target yolov8 \
  --ontology '{"milk bottle": "bottle"}' \
  --output ./dataset \
  --epochs 200 \
  --upload-to-roboflow false
```

Before running a full command, verify the input image folder, ontology JSON, selected plugin packages, target model training requirements, and output directory. Avoid `-y true` until plugin installation side effects are approved.

## Important Snapshot Caveats

- `--models`, `--upload-to-roboflow`, `-y`, and `--test` are Click boolean options that expect explicit values in this snapshot's help output, for example `--upload-to-roboflow false`.
- Source verification shows `SUPPORTED_MODEL_TYPES` is `['detection', 'segmentationclassification']` because `"segmentation" "classification"` is missing a comma. The docs mention detection, segmentation, and classification, but `--model_type classification` may fail in this snapshot.
- `import_requisite_module` may call `pip install autodistill_<alias>` when noninteractive install is enabled. That package naming does not match every published plugin's hyphenated distribution spelling in documentation, so inspect before relying on auto-install.

## When to Route Elsewhere

- To validate the dataset files generated by the CLI, use [dataset labeling](../dataset-labeling/SKILL.md).
- To implement or debug a plugin class, use [ontologies and model interfaces](../ontologies-and-model-interfaces/SKILL.md).
- To debug image input conversion or visualization, use [utilities](../utilities/SKILL.md).