SASS-Bench — Schema-Abstention Stress Suite

Data and scored results for the paper "Valid JSON and False Facts under Schema Pressure in Tiny Language Models."

Structured-output pipelines often treat parseable JSON as a proxy for reliability. SASS-Bench isolates a subtler failure in sub-billion-parameter language models: a required schema slot can be filled with a fabricated value even when the source text does not support it — although the schema permits null for every field. We call the effect schema pressure and quantify it with the Schema Pressure Index (SPI).

SPI-C2 = FalseFillRate(strict JSON) − FalseFillRate(free text) Positive SPI ⇒ imposing a JSON schema increased fabrication of unsupported fields.

• Code / reproduction harness: https://github.com/MohammadNuramin/sass-bench
• Benchmark size: 1,000 synthetic items (seed 42), 5 evidence regimes
• Main run: 10 sub-1B models × 6 conditions (+ int8/int4 quantized variants) = 109,000 generations
• Decoding: deterministic (temp 0) for the main grid; a temp-0.2 ×3 replication is included

What's in here

benchmark/
  sass_bench.jsonl        the 1,000-item benchmark: context, target entity, gold object, evidence status
  person_event_v1.schema.json   the flat 5-field extraction schema
main/                     primary deterministic run (C0–C5, incl. int8/int4 variants)
  raw_outputs.jsonl       109,000 raw generations (one record per model×condition×item)
  field_metrics.csv       per-field scoring (the atomic scored table)
  summary_by_model_condition.csv
  schema_pressure_index.csv
  spi_confidence_intervals.csv    paired cluster bootstrap, B=10,000, 95% CIs
  summary_by_challenge.csv
  qualitative_failures.csv
  experiment_config.json  hardware / software versions / model revisions
sensitivity_c0_json_fallback/     C0 re-scored with a JSON-aware baseline (fairness check for extraction-tuned models)
robustness_temp0.2_c0c3/          temp-0.2, top_p 0.95, ×3 replication of C0 and C3
robustness_temp0.2_c2/            temp-0.2, top_p 0.95, ×3 replication of C2

Conditions

	prompt
C0	free-text extraction (baseline)
C1	loose JSON (no strictness)
C2	strict JSON (exact keys, no explanation)
C3	strict JSON + explicit null rule
C4	strict JSON + few-shot null demonstrations
C5	schema-constrained decoding

Evidence regimes (200 items each)

supported · missing · ambiguous · contradictory · distractor. Every non-supported item has exactly one controlled stress field whose gold value is null (or, for distractor, whose gold value is the target entity's true value while a wrong value appears nearby); the other fields are supported.

Record schemas

benchmark/sass_bench.jsonl id, domain, challenge_type, target_entity, context, schema_name, gold, field_status, evidence_spans, distractors

main/raw_outputs.jsonl run_id, model, model_revision, model_parameter_count, precision, condition, temperature, top_p, max_new_tokens, example_id, prompt, raw_output, parsed_output, parse_success, schema_valid, latency_ms, tokens_in, tokens_out, used_chat (parsed_output is intentionally left null in raw outputs — the scorer is the single source of parsing truth.)

*/field_metrics.csv (one row per scored field) run_id, model, condition, example_id, challenge_type, field, field_status, gold_value, predicted_value, gold_norm, pred_norm, is_correct, is_false_fill, is_correct_null, predicted_null, is_distractor_error, parse_success, schema_valid

*/summary_by_model_condition.csv model, condition, n_outputs, json_parse_rate, schema_validity_rate, supported_accuracy, false_fill_rate, null_recall, abstention_precision, distractor_error_rate, field_faithfulness_rate, format_truth_gap

*/schema_pressure_index.csv model, ffr_c0, ffr_c2, spi_c2, ffr_c3, spi_c3, ffr_c4, few_shot_improvement_c3_minus_c4

Headline result

Strict JSON significantly increases fabrication in the models capable of emitting valid JSON:

Model	FalseFill C0→C2	SPI-C2 [95% CI]
Qwen3-0.6B	0.22 → 0.70	+0.485 [0.44, 0.52]
Qwen3.5-0.8B	0.29 → 0.71	+0.422 [0.38, 0.46]
Granite-4.0-H-350M	0.39 → 0.65	+0.255 [0.22, 0.29]
Granite-4.0-350M	0.62 → 0.81	+0.192 [0.16, 0.22]
LFM2.5-350M	0.72 → 0.76	+0.045 [0.005, 0.085]

The effect is conditional, not universal: some models already false-fill in free text, and others avoid false-fills only by failing to emit valid JSON. Constrained decoding (C5) gives 1.00 parse/schema validity for every compatible model while false-fill rates stay as high as 0.73. Explicit null rules mitigate more reliably than few-shot examples.

Loading

from datasets import load_dataset

bench = load_dataset("nuramin/sass-bench", "benchmark", split="train")
spi   = load_dataset("nuramin/sass-bench", "schema_pressure_index", split="train")

import pandas as pd
fm = pd.read_csv("hf://datasets/nuramin/sass-bench/main/field_metrics.csv")
false_fill = fm[fm.is_false_fill == 1]

Scoring notes

Scoring is deliberately conservative: a plausible guess is counted wrong unless directly supported by the context. Model outputs are not repaired before scoring, and the benchmark was frozen before any outputs were seen. See the harness for the exact scorer.

Citation

A full citation will be added on publication. Until then, please cite this dataset and the accompanying repository (https://github.com/MohammadNuramin/sass-bench). Author and affiliation in the paper are anonymized pending review.

License

Released under CC-BY-4.0. The benchmark items are synthetic (fictional contexts, no personal data); model generations are provided verbatim for research reproducibility.