Tasks

Tasks transform artefacts. Choose a task class in pipeline.yaml, give it a name, and connect its input and output artefact nodes. Commands use the configured name, not the class name:

collectra --pipeline pipeline.yaml task --help
collectra --pipeline pipeline.yaml task detect_lines --help
collectra --pipeline pipeline.yaml task detect_lines run --help

• Object Detection
  • ObjectDetectionYOLO
  • ObjectDetectionRFDETR
  • ObjectDetectionDETR
• Image segmentation
  • Installation
  • Mask-supervised training
  • Metrics, W&B and model selection
  • Troubleshooting
• Line Detection
  • LineDetectorRLSA
  • LineDetectorSurya
• Image Classification
  • ImageClassifierYOLO
  • ImageClassifierTorchvision
  • ImageClassifierHuggingFace
• OCR
  • OCRSuraya
  • Commands
• LLM
  • Commands
• Entity Resolution
  • LLMCanonicaliser
  • IRNResolver
• Image Orientation
  • ImageOrienter
  • Commands
• Text Processing
  • ConcatenateText
  • Commands

Command overview

Built-in task commands

Task class	Commands
ObjectDetectionYOLO	run, train
ObjectDetectionRFDETR	run, train
ObjectDetectionDETR	run, train
ImageSegmentationSam3	run, train
LineDetectorRLSA	run
LineDetectorSurya	run
ImageClassifierYOLO	run, train
ImageClassifierTorchvision	run, train
ImageClassifierHuggingFace	run, train
OCRSuraya	run
LLM	run
LLMCanonicaliser	run
IRNResolver	run
ImageOrienter	run
ConcatenateText	run

Running tasks

Every task supports run INPUTS... and run --help. Shared options are --output, --verbose, --usage, and --render. --usage records LLM token usage when supported, and --render requires Graphviz. Task-level run executes only the selected task. Its input artefacts must be present in the supplied results, or be the root image supplied as a raw file.

To execute a task and then its downstream tasks, use the pipeline command:

collectra --pipeline pipeline.yaml run results/page001.collectra --task detect_lines

Training tasks

Only the detection and classification tasks listed above expose train. Training consumes annotated Collectra result directories. Options differ by backend; use the particular task’s train --help. Torchvision and Hugging Face classifiers support opt-in experiment logging with --wandb. RF-DETR exposes --no-wandb to disable its default logging.

Training updates the task’s model in the pipeline and retains training logs by default. See Quickstart for preparing annotations and a complete training example. Evaluation and image extraction are artefact commands, not task commands.