Set

Use Set for multi-label discrete values or targets from a bounded, learned vocabulary: tags, permissions, flags, detected concepts, applicable categories, and similar multi-label fields.

{
  "tags": ["vip", "international", "vip"]
}

import json2vec as jv

tags = jv.Set(
    "tags",
    size=2048,
    p_unavailable=0.02,
    threshold=0.75,
)

tags

tags [set] active
 pooling=query weight=1 p_mask=0 p_prune=0 n_heads=4 n_linear=1
 size=2048 p_unavailable=0.02 threshold=0.75

The input to a Set and a Branch of Category values can look similar, but use Set("tags") for unordered labels. An Branch of Category expects order or item-level structure to matter.

Input Values

Set accepts common multi-label shapes:

• None becomes a null field state.
• A string becomes a single-item set.
• An iterable becomes a multi-item set.
• A non-iterable scalar becomes a single-item set.
• An empty iterable becomes a valued field with no active labels.

Repeated labels collapse into one active vocabulary entry. The vocabulary is learned online from training data and saved with the model.

Examples

Common set fields include:

• Product tags, content topics, user interests, or applicable policies.
• Permissions, feature flags, alert codes, or rule hits.
• Detected entities or concepts from an upstream parser.
• Multi-label outcomes where more than one class can be true at once.

Use Branch(tag=Category()) instead when repeated labels need positions, attributes, or their own local context.

Configuration

Option	Default	Notes
size	10000	Maximum number of learned labels. One extra internal bucket is reserved for unavailable labels.
p_unavailable	0.01	During training, randomly routes known labels to the unavailable bucket so the decoder learns that case.
threshold	None	Optional prediction-time probability threshold. When set, prediction output only includes labels at or above the threshold.

Shared Preprocessing State

Set maintains an online vocabulary shared by encoding workers and saved with the model. Training data can add new labels until size is full. Validation, test, and prediction data do not expand the vocabulary.

Unknown labels after training are represented internally by a reserved unavailable slot at index size. Prediction output only includes probabilities for labels in the learned vocabulary.

p_unavailable randomly removes some known training labels from their normal vocabulary slots and marks the unavailable slot for that set value. This teaches the model that a valued set can contain labels that are not individually known.

Target Behavior

When a Set is masked or used as a target, the decoder predicts:

• state: probabilities for valued, null, padded, and masked.
• content: independent logits for each learned vocabulary item.

Content loss uses binary cross entropy because each label is independently present or absent.

Prediction Output

Without threshold, Model.predict(...) returns state probabilities and one probability array per known label:

{
    "record/tags": {
        "state": {"valued": ..., "null": ..., "padded": ..., "masked": ...},
        "content": {
            "vip": ...,
            "international": ...,
        },
    }
}

{"record/tags": {"state": {"valued": "text/plain:Ellipsis", "null": "text/plain:Ellipsis", "padded": "text/plain:Ellipsis", "masked": "text/plain:Ellipsis"}, "content": {"vip": "text/plain:Ellipsis", "international": "text/plain:Ellipsis"}}}

With threshold=0.75, content is a nested JSON-like structure that only includes labels whose probabilities meet the threshold:

{
    "record/tags": {
        "state": {"valued": ..., "null": ..., "padded": ..., "masked": ...},
        "content": {"vip": 0.91},
    }
}

{"record/tags": {"state": {"valued": "text/plain:Ellipsis", "null": "text/plain:Ellipsis", "padded": "text/plain:Ellipsis", "masked": "text/plain:Ellipsis"}, "content": {"vip": "text/plain+float:0.91"}}}

Notes

Unknown labels at validation, test, or inference time are represented internally with the reserved unavailable bucket. Prediction output only includes labels in the learned vocabulary. Thresholding is an inference-output filter; it does not change training loss.