---

name: data-labeling-and-training-datasets description: "Frame extraction, label checks, training dataset and shuffle creation, data conversion, and tiny DeepLabCut-style fixtures." disable-model-invocation: true metadata: disco-role: operating license: LGPL 3.0

Data Labeling And Training Datasets

Use this sub-skill when the current DeepLabCut task is about turning project videos and annotations into valid DeepLabCut labeled-data and training-dataset artifacts.

Covers

• Extracting frames with deeplabcut.extract_frames for existing projects.
• Checking labels with deeplabcut.check_labels after GUI labeling, external import, or conversion.
• Understanding and repairing DeepLabCut labeled-data/ annotation tables for standard and multi-animal projects.
• Creating shuffles and train/test splits with create_training_dataset, create_training_dataset_from_existing_split, create_training_model_comparison, and mergeandsplit.
• Converting annotation files with convertcsv2h5 and converting older single-animal labels into maDLC-style labels with convert2_maDLC.
• Creating tiny safe standard or multi-animal fixture projects with the bundled script.

Route elsewhere

• New-project creation, installation, package imports, environment preparation, and initial config.yaml project setup: use install-and-project-setup.
• Training, evaluation, model selection for real training runs, inference, analyzing videos, and exported model runtime: use pytorch-training-evaluation-inference.
• Tracklet conversion, stitching, identity tracking, and multi-animal temporal assembly after inference: use multi-animal-tracking.
• Filtering predictions, outlier-frame refinement, 3D workflows, labeled-video export, trajectory plots, and other post-prediction outputs: use postprocessing-3d-video-exports.

Read in order

1. Data formats for the exact folder/table layouts and trainset metadata expected by DeepLabCut.
2. Workflows for API sequences, safe arguments, and boundary decisions.
3. Troubleshooting when labels are not found, shuffles collide, CSV/HDF conversion fails, or multi-animal columns are malformed.
4. Tiny fixture script when a small local standard or multi-animal project is needed for format or routing checks without GUI, training, downloads, or network access.

Operating reminders

• Treat config.yaml as the authority for scorer, project path, videos, body parts, individuals, training fractions, iteration, and engine.
• Do not start training from this sub-skill. Stop after label validation and dataset/shuffle creation, then route to training/evaluation.
• In non-interactive automation, avoid prompts by setting userfeedback=False only after confirming that the target output may be overwritten or is disposable.
• Keep annotation paths portable: rows should identify images under labeled-data/<video-stem>/..., and the project path in config.yaml should point at the project root on the machine where the operation is being run.