Unseen1980/mockgen-classifier-data-v1

Schema-property semantic classification data (v1)

Field metadata — property name, label, containing entity, neighbouring property names, type, annotations — paired with a semantic hint such as country, currency, person_full_name or gl_account. Each row describes a schema field, never a business record. No values are included.

The task: given only what a schema says about a field, decide what the field means, so a mock-data generator can produce something plausible for it.

Files and what is actually in them

Read the `unknown` column before using this. public-real.jsonl is the bulk of the corpus and 86% of it is unlabelled — those rows are real field metadata harvested from public API descriptions (apis-guru) for which no semantic hint was assigned. They are useful as unlabelled or negative data, and as a realistic sample of what enterprise schema fields actually look like, but they are not 30k labelled examples.

train.jsonl and validation.jsonl are small project-authored fixture sets (135/15 rows), not a large training split. A previous version of this card implied otherwise.

Rows previously marked REVIEW_ME — an internal work-in-progress marker that should not have been published — have been removed (2,120 rows from public-real.jsonl).

Row schema

• —property_name — schema property identifier
• —label — human-readable label where available
• —entity_type_name — containing entity or schema object
• —neighbor_properties — sibling property names, the main contextual signal
• —type — declared type (Edm.String, string, …)
• —sap_annotations_summary — comma-separated annotation terms where present
• —locale, source_service, source_kind
• —hint — the semantic label, or unknown

Label space

52 hints, covering the 50 used by the MockGen runtime classifier plus booking_id, travel_id and work_center.

Provenance

public-real.jsonl comes from the public apis-guru OpenAPI directory. The remaining files are project-authored or reviewed public OData fixtures. Excluded by construction: private metadata, service keys, generated rows, and business values.

Honest baseline

On a 300-row human-adjudicated held-out benchmark, the classifier trained from this data reaches ~38% accuracy / 0.29 macro-F1. Reported in-distribution validation accuracy on synthetic splits runs far higher and does not transfer — the held-out number is the one to compare against.

License

Apache-2.0.