SchemaStress

SchemaStress is a controlled synthetic benchmark for structured output reliability under schema constraints.

Dataset Summary

SchemaStress evaluates model behavior on schema-bounded structured tasks:

• prompt-to-structured generation
• invalid-candidate repair
• validity reasoning support via error tags and paths

The benchmark focuses on outputs that must be parseable, schema-valid, and semantically coherent.

Supported Configs

• form_easy: reimbursement record objects
• workflow_hard: nested approval workflow objects with cross-field semantics

Data Splits

Each config provides:

• train
• validation
• test_in_domain
• test_ood

Example file pattern:

• data/<config>/<split>-00000.jsonl

Data Fields

Primary fields:

• instruction
• schema_json
• semantic_constraints
• target_object
• candidate_object
• validity_label
• error_tags
• error_paths
• split

Full dictionary: docs/FIELD_DICTIONARY.md

Dataset Creation

Generation is synthetic and deterministic by seed.

Build command:

make highvar

Default build in this repository generates 5000 total rows (2500 per config).

Validation and QA

Quality checks include:

• schema + semantic validation of targets
• split-policy consistency checks
• repair-label and error-annotation consistency checks

Outputs:

• reports/phase4/<config>_qa_report.json
• reports/diversity_report.json

Baselines

Deterministic baselines are provided for:

• generation validity
• repair success
• candidate validity classification

Run:

make baselines

Output:

• reports/phase5_baselines.json

Intended Uses

• benchmarking structured output reliability
• evaluating repair pipelines
• measuring robustness to schema complexity and OOD composition

Out-of-Scope Uses

• factual knowledge benchmarking
• direct use as real production business records

Limitations

• synthetic language may retain template artifacts
• schema subset does not cover full industrial JSON Schema complexity
• benchmark semantics are controlled and may not capture all edge cases

Citation

If you use this dataset, cite the repository and include the release manifest seed/config details from metadata/build_manifest.json.