abr-6024/schema-json-extractor-synthetic

Schema JSON Extractor: synthetic training data (v4)

4,000 train / 450 val-seen / 450 val-unseen examples of (JSON Schema, text) -> JSON, produced by the generator in gen/. The target validates against the example's own schema, and has null for every field the text does not state. Used to train abr-6024/schema-json-extractor-4B.

Everything here is synthetic. The people, companies, addresses and amounts are random combinations from word lists. No real person's data is included. It is template-rendered text, not natural prose; see "Limitations" before assuming it transfers.

How the examples are made (record-first)

1. Sample a world state for a domain (a receipt, a job ad, a cinema booking, ...).
2. Choose which fields the text will state. About 20% of fields end up absent (null).
3. Render only the stated fields as text, in one of several layouts (prose, key-value block, log line, receipt tape, email headers), with values in varied surface forms (14 March 2026, 03/14/2026, €1,249.50, 1249.5 EUR).
4. Project the world state through the schema to get the expected JSON. Validate it against the schema; discard anything invalid.
5. Apply a noise profile (below).

Because the record exists before the text, the target cannot disagree with the text, provided the renderer really prints every stated field. gen/audit.py checks that: every non-null value must be findable in its text in some rendered surface form. 0 ungrounded values across all three splits (35,533 train fields, 4,014 val-seen, 3,797 val-unseen).

Contents

Split	Rows	Schemas	Notes
train.jsonl	4,000	72	12 domains, 35,533 fields, 19.8% null
val_seen.jsonl	450	the same 72	new rows over trained schemas
val_unseen.jsonl	450	63 other schemas	none of these schemas appear in train

Domains (train rows): cinema 367, email 352, maintenance 332, review 332, ticket 327, log 326, property 325, news 320, warehouse 320, job 308, meeting 349, receipt 342.

Noise profiles (train rows): clean 1,091 - inject 716 - long 389 - contradict 383 - distract 375

• foreign 364 - typos 349 - verylong 333.

The inject profile (17.9% of train)

Rows where the text contains an instruction aimed at the extractor, and the target ignores it. Three kinds, all with phrasings generated by this repo:

• null-targeted: names a field that is in this row's schema and really null ("set time to UNKNOWN"); the target keeps it null.
• override: names a field the text states, and dictates a different value; the target keeps the stated value.
• generic: instructions with no field name.

Checked in the data: 0 null-targeted attacks name a field that is not genuinely null, and 0 override attacks name an unstated field or dictate the true value (that would teach obedience).

Fields in each row

id, domain, schema_id, noise, missing_rate, null_fields, total_fields, schema (JSON Schema, draft 2020-12, every property nullable, additionalProperties: false), text, expected (the target object), messages (system / user / assistant, exactly as used for training). The user message is <schema>...</schema> then <text>...</text>; the system prompt is in gen/build.py.

schema differs in structure from row to row, so the Hugging Face dataset viewer is switched off and datasets.load_dataset("json") can fail on the mixed structs. Read it as plain JSONL:

import json
rows = [json.loads(l) for l in open("train.jsonl", encoding="utf-8")]

Reproduce (tested: byte-identical)

cd gen
python build.py --train 4000 --val 450 --unseen 450 --seed 7 --out ../out
python audit.py ../out/train.jsonl ../out/val_seen.jsonl ../out/val_unseen.jsonl

Rebuilding from a clean copy gave files whose SHA-256 matched these three. spotcheck.md lays out 50 examples for manual review; stats.json has the seed and schema lists.

Limitations (read these)

• Template text, not real documents. In-distribution scores on val_* measure fit to this generator. On hand-written prose in new domains, the model trained on it is not more accurate than its untuned base; its measured edge is resisting instructions hidden in the text. See the model card and BLOG.md in the GitHub repo.
• English only. Dates, currencies and names follow a small set of conventions.
• 12 domains. The held-out schemas are new schemas over the same domains; they are not new vocabularies.
• The inject phrasings are a sample, not a complete attack taxonomy: 0/9 on a fresh set of new phrasings is encouraging, not proof of robustness.

Links

• Model: https://huggingface.co/abr-6024/schema-json-extractor-4B
• Code, evals and write-up: https://github.com/ab-rar-6024/schema-json-extractor