Category

Use Category for a single label from a bounded vocabulary: class labels, country codes, product families, merchant IDs, and similar discrete values.

{
  "merchant": "ACME_MARKET",
  "merchant_id": "12345"
}

import json2vec as jv

merchant = jv.Category(
    "merchant",
    size=4096,
    topk=[3, 10],
    p_unavailable=0.02,
)

merchant

merchant [category] active
 pooling=query weight=1 p_mask=0 p_prune=0 n_heads=4 n_linear=1
 size=4096 p_unavailable=0.02 topk=[3, 10]

Use Category when exactly one label is present. Use Set when the field can contain multiple labels, and use Entity when the goal is local equality matching inside repeated records rather than a persistent vocabulary.

Input Values

Category expects scalar labels. Strings are the normal input. Use a preprocessor if upstream labels or codes need to be converted into stable strings or otherwise cleaned. None is encoded as a null state, and missing branch positions are encoded as padded state.

Examples

Common category fields include:

• A supervised class label such as fraud, churn, species, or diagnosis.
• A bounded business label such as customer tier, region, product family, or channel.
• A stable categorical code such as country, currency, device type, or merchant category.

Avoid using Category for massive identifiers such as customer IDs, account IDs, session IDs, or device IDs. A large global vocabulary is memory-heavy and usually generalizes poorly. Keep unique IDs as metadata, or use Entity when the model only needs local equality within a repeated context.

target = jv.Category("fraud", target=True, size=3, topk=[2])

target

fraud [category] active target
 pooling=query weight=1 p_mask=0 p_prune=1 n_heads=4 n_linear=1
 size=3 p_unavailable=0.01 topk=[2]

Configuration

Option	Default	Notes
size	1024	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.
topk	[]	Optional top-k accuracy metrics and prediction candidates. Values must be positive, not 1, and less than size.

Shared Preprocessing State

Category 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.

Labels that are not in the learned vocabulary keep state valued, but their content is routed to a reserved unavailable bucket at index size. This preserves the fact that the field existed while avoiding unstable vocabulary growth after training.

p_unavailable deliberately complicates training by randomly routing some known training labels into that unavailable bucket. This gives the decoder examples of valued-but-unavailable content before it sees truly unseen labels during validation or inference.

p_unavailable gives the decoder training examples for labels that are present but outside the learned vocabulary, which is the serving-time behavior for unseen labels.

Target Behavior

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

• state: probabilities for valued, null, padded, and masked.
• content: a categorical distribution over learned vocabulary labels plus the internal unavailable bucket.

Top-k metrics are tracked for each configured value in topk.

Top-K Nuances

topk serves two roles:

• During training and evaluation, json2vec tracks a separate top-k accuracy metric for each configured value.
• During prediction, the output contains one candidate list sized to max(topk), capped by the current learned vocabulary size.

topk=1 is disallowed because top-1 is already represented by the primary content.value and content.probability fields. Values must be less than size so the requested candidates stay inside the learned label space. Duplicate values are normalized away.

The reserved unavailable bucket can affect training loss and metrics, but it is not emitted as a prediction candidate.

Prediction Output

Rendered examples use string paths for readability. The Python dictionary returned by Model.predict(...) is keyed by jv.Address objects.

Model.predict(...) returns the most likely known label, its probability, and the candidate list requested by topk:

{
    "record/merchant": {
        "state": {"valued": ..., "null": ..., "padded": ..., "masked": ...},
        "content": {
            "value": ...,
            "probability": ...,
            "topk": ...,
        },
    }
}

{"record/merchant": {"state": {"valued": "text/plain:Ellipsis", "null": "text/plain:Ellipsis", "padded": "text/plain:Ellipsis", "masked": "text/plain:Ellipsis"}, "content": {"value": "text/plain:Ellipsis", "probability": "text/plain:Ellipsis", "topk": "text/plain:Ellipsis"}}}

The special <unavailable> token is not emitted as a predicted label.

Notes

Use Set when a field can contain multiple labels (such that categories are not mutually exclusive). Use Entity when the goal is local identity matching inside repeated records rather than a persistent global vocabulary. Use Text if you want to break category labels down into word pieces.