---

name: dataset-labeling description: "Guides Autodistill dataset auto-labeling workflows, output layouts, and safe dummy verification for detection and classification base models." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

Autodistill Dataset Labeling

Use this sub-skill when a task asks how to auto-label image folders, inspect or validate Autodistill dataset output, choose CaptionOntology prompts/classes for labeling, use SAHI/NMS during labeling, or debug core .label() behavior without running a real foundation model.

Autodistill's core package supplies base-model interfaces and dataset writers. Real foundation models such as Grounding DINO, Grounded SAM, CLIP, YOLOv8, or cloud plugins live in separate autodistill-* packages and may require downloads, credentials, GPU, or long training. For plugin aliases and CLI orchestration, read the CLI and model registry sub-skill. For implementing a custom model class, read ontologies and model interfaces.

Quick Route

• Programmatic image auto-labeling: read workflows for .label() patterns, ontology setup, SAHI/NMS switches, and Roboflow upload boundaries.
• Detection/classification output layout: read data formats before validating data.yaml, YOLO labels, train/valid splits, or classification folders.
• Signatures and defaults: read API reference for source-verified method names, parameters, return types, and enum values.
• Problems while labeling: read troubleshooting for empty inputs, extension mismatches, missing confidence values, memory pressure, plugin/backend failures, and credentialed Roboflow paths.
• Safe local proof of core dataset writing: run scripts/create_tiny_detection_dataset.py to create a tiny deterministic YOLO-style detection dataset without any model plugin.

Core Concepts

1. Build an ontology, usually CaptionOntology({"prompt sent to base model": "saved class name"}).
2. Instantiate a base model plugin or a custom subclass whose task matches the dataset you want.
3. Call base_model.label(input_folder=..., extension=".jpg", output_folder=...).
4. Validate the generated dataset layout before handing it to a target model plugin.

In this source snapshot the verified source method is label(). Some public docs mention label_folder(); treat that as a stale alias unless a concrete plugin adds it. The verified SAHI method on DetectionBaseModel is sahi_predict(), while .label(..., sahi=True) uses supervision.InferenceSlicer internally.

Minimal Programmatic Pattern

from autodistill.detection import CaptionOntology
from autodistill_grounding_dino import GroundingDINO

ontology = CaptionOntology({"milk bottle": "bottle", "bottle cap": "cap"})
base_model = GroundingDINO(ontology=ontology)
base_model.label(input_folder="images", extension=".jpg", output_folder="dataset")

This pattern proves the core call shape. The plugin package, model weights, hardware, and optional training target must be selected and verified separately.

Validation Checklist

• ontology.classes() contains exactly the labels expected by the target model.
• The input_folder contains files matching extension; default is .jpg only.
• Detection outputs contain confidence if record_confidence=True.
• The generated detection dataset has data.yaml, train/valid image folders, and train/valid label folders.
• Any full model run has explicit approval for downloads, credentials, GPU, or long training.

Safe Smoke Script

Run the bundled dummy script after installing autodistill and supervision:

python scripts/create_tiny_detection_dataset.py --keep

It uses a deterministic in-process DetectionBaseModel subclass, creates two tiny images, calls the inherited label() implementation, and checks the output layout. Passing this script means the core dataset writer works; it does not verify any external base-model plugin.