Artefact Types

Artefacts are the data flowing through a pipeline: source images, crops, text, and class labels. results.yaml stores their types, IDs, data, and parent relationships. Tasks produce artefacts; artefact commands operate on saved results, using the configured node name:

collectra --pipeline pipeline.yaml artefact --help
collectra --pipeline pipeline.yaml artefact lines --help

For example, lines names a node of ImageCrop artefacts, rather than an individual crop ID or the class name ImageCrop.

• Image
• ImageCrop
• ImageSegmentation
• Text
• Link
• Custom artefact displays

Command overview

Built-in artefact commands

Type	Commands	Behaviour
Image	extract, evaluate	Extraction is supported. The inherited evaluation command has no comparison implementation for whole Image artefacts.
ImageCrop	extract, evaluate	Export crop images and evaluate box intersection-over-union.
Text	evaluate	Compare text using sequence similarity.
Link	evaluate	Evaluate references; see the Link page for comparison behaviour.

Evaluation

evaluate PREDICTED_FOLDER GOLD_FOLDER compares matching result directories for the selected node. Shared options are:

• --threshold: minimum matching score, default 0.5.

• --match-by-order: use positional matching instead of the normal matching.

• --output: write evaluation metrics to a CSV file.

Reports are also printed to the terminal. For extraction, pass individual result directories or a shell-expanded glob; extract does not scan an arbitrary parent directory for results. Every command supports --help.

Artefacts do not expose run or train; those are task commands. See GUI to inspect and edit result artefacts visually.