Model categorization taxonomy
taxonomies/model-taxonomy.yaml is the canonical source of searchable model
categories. It owns stable machine IDs, display labels, architecture paradigms,
and model-type compatibility rules. JSON Schema validates the shape of category
references; the model-registry semantic validator resolves those references
against this file.
Do not copy sanctioned lists into model definitions, browser code, API code, or database migrations. Each model declares references:
model_type: vision
tasks: [image_classification]
modalities:
inputs: [image]
outputs: [class_scores]
architecture:
id: resnet
The facets are intentionally independent:
- model type describes the broad product class, such as Vision, LLM, or VLM;
- task describes what the model does;
- modality describes semantic inputs and outputs;
architecture.idnames the model family, the upstream model line without its size:yolo26_det, notyolo26n_det, andresnext, notresnext_50. The model'sidanddisplay_namecarry the size. The browser labels this filter Model family;- the family's
paradigm(CNN or Transformer) is derived, not declared. The browser labels it Architecture.
To categorize a model, take every ID from the taxonomy. When none fits, add the
entry to taxonomies/model-taxonomy.yaml in the same pull request.
Each architecture entry requires a short description for the family card,
alongside its existing display_name, paradigm and allowed_model_types:
architectures:
yolov8_obb:
display_name: YOLOv8-obb
description: Detects objects with rotated bounding boxes in aerial imagery.
paradigm: cnn
allowed_model_types: [vision]
Describe the whole family, without size-specific claims. Keep this text in the
taxonomy; individual model.yaml files do not repeat it. The browser derives
family membership, tasks and the representative hero from available models.
Cards group and sort by the existing architecture ID and use display_name
as their title. Detection, segmentation, pose and OBB keep their separate IDs.
Publication identity
The taxonomy has no release version of its own. CI publishes it with every catalog, so each published catalog keeps the exact taxonomy that describes its models.
No conversion, compile, or benchmark fingerprint includes the taxonomy. A taxonomy change, such as a new family description, therefore reruns nothing, and existing builds and results stay reusable.
All authoritative category facets are mandatory. Validation rejects a missing field, an unknown sanctioned ID, or an incompatible model type before CI compiles anything.
Renames and removals
Canonical IDs are stable. To rename or remove a concept, change its taxonomy entry and every model that references it in one pull request; there is no alias, deprecated entry, or taxonomy release number. Full-catalog validation rejects a model that still references a removed ID before publication.
To remove a model or a build profile, delete it from the repository. After the next catalog publication, the model browser no longer lists it.