StructFix-Bench

A benchmark for schema-aware structured output recovery — repairing broken outputs from agents, tool calls, and LLM workflows.

What it tests

Unlike JSON syntax repair benchmarks, StructFix-Bench focuses on semantic and schema-level recovery:

• Replacing invalid enum values with valid ones
• Adding missing required fields
• Correcting type mismatches
• Recovering tool call arguments from Python syntax
• Reconstructing outputs from truncated agent chains
• Extracting JSON from markdown/text wrappers

Dataset splits

Split	Examples	Field names	Enums
train	200,000	Mixed (semantic + shuffled)	Mixed (realistic + synthetic)
validation	20,000	Semantic	Realistic
test_seen	10,000	Semantic	Realistic
test_unseen	10,000	Semantic	Realistic
test_random_fields	10,000	Random hex (f_8f31a7)	Synthetic

Corruption types (28 total)

Syntactic (10)

single_quotes, unquoted_keys, trailing_comma, markdown_fences, extra_text_before, extra_text_after, truncated_object, truncated_array, missing_bracket, missing_brace

Semantic (9)

wrong_type, invalid_enum, missing_required, null_required, extra_field, boolean_as_string, number_with_unit, synonym_enum, date_format_error

LLM-like (3)

markdown_explanation, partial_structure, thought_process

Tool/Agent (6)

tool_call_bad_format, tool_call_text_mix, tool_call_python_syntax, tool_call_partial_args, tool_call_wrong_param, agent_chain

Input format

Each example uses ConstraintDSL to represent the schema:

TASK repair_structured_output

SPEC
FIELD priority TYPE string VALUES low|medium|high REQUIRED yes
FIELD description TYPE string REQUIRED yes

BROKEN_OUTPUT
{"priority":"urgent"}

Target: {"priority":"high","description":""}

Key results

Method	Schema Success (unseen)	Schema Success (random fields)
json-repair	65.2%	64.0%
CodeT5+ + JSON Schema	55.0%	—
CodeT5+ + ConstraintDSL	96.3%	91.9%

Schema ablation study

Full ablation results are included in results/all_results.json:

DSL variant	Schema Success
Full ConstraintDSL	96.3%
No DSL at all	78.4%
Dummy DSL	75.0%
Conflicting DSL	88.3%
Shuffled field names	73.4%
Shuffled enum values	89.8%
Shuffled required flags	93.2%
Minimal DSL	82.5%
Field names only	72.2%

Known limitations

The benchmark uses synthetically generated schemas and corruptions. A showcase validation with 25 hand-crafted real-world examples revealed that model performance degrades when field names overlap with training vocabulary — the model may substitute field names (e.g., action → active, contract_id → consign_id). See the model card for details.

Files

• train.jsonl / validation.jsonl / test_seen.jsonl / test_unseen.jsonl / test_random_fields.jsonl
• results/all_results.json — all experiment results including ablations

Each example contains

{
  "input": "TASK repair_structured_output\n\nSPEC\n...",
  "target": "{\"priority\":\"high\",\"description\":\"\"}",
  "corruption_type": "invalid_enum",
  "schema": {"type": "object", "properties": {...}, "required": [...]},
  "invalid_json": "{\"priority\":\"urgent\"}",
  "error": "Field 'priority' has invalid enum value 'urgent'",
  "target_json": "{\"priority\":\"high\",\"description\":\"\"}"
}

Citation

@software{structfix_bench,
  title = {StructFix-Bench: A Benchmark for Schema-Aware Structured Output Recovery},
  author = {Ottema},
  year = {2026},
  url = {https://huggingface.co/datasets/ottema/structfix-bench}
}

License

Apache-2.0