docs/FIELD_DICTIONARY.md

SchemaStress Field Dictionary

This document defines the row-level fields used in the generated JSONL artifacts under data/release/.

Core Identity and Routing

• id: unique row id (schemastress_<config>_<task>_<index>)
• benchmark_family: constant family label (schemastress)
• config: benchmark config (form_easy, workflow_hard)
• task: task type (generation, repair)
• split: data split (train, validation, test_in_domain, test_ood)
• seed: generation seed for reproducibility

Scenario and Difficulty Metadata

• domain: domain label (reimbursement_request, approval_workflow)
• difficulty: coarse difficulty bucket for config (easy, hard)
• grammar_depth: approximate structural depth indicator
• num_fields: number of top-level fields in target_object
• num_required_fields: number of required fields in schema_json
• num_constraints: number of semantic constraints for the config
• object_token_length: tokenized length proxy for target_object
• instruction_token_length: tokenized length proxy for instruction

Task Inputs

• instruction: natural language request to satisfy
• schema_id: schema registry id for the row
• schema_json: JSON Schema-like object used for structural validation
• semantic_constraints: benchmark-side semantic predicates

Task Targets and Candidates

• target_object: valid gold object for the instruction and schema
• candidate_object: candidate object (present for repair, null for generation)
• validity_label: validity label for candidate_object (invalid in current repair generation)

Error and Repair Annotations

• error_tags: set of validator error categories for candidate_object
• error_paths: failing field paths (JSONPath-like)
• repair_distance: proxy for number of failing paths

Notes

• For generation rows, candidate_object and validity_label are null.
• For repair rows, candidate_object is intentionally corrupted and expected to be invalid before repair.