balys/extract-index-suite

SystemOne Extract Index: suite 0.1

The frozen test suite of the SystemOne Extract Index. It measures one-pass typed extraction: a model reads a document (state) and a map of typed fields, and returns a value or null for each field. The kit, scorer and rules are at https://github.com/balys/extract-index.

Test	Area	Requests	Fields	Field types	Gold null	Licence	Source
squad2	Spans	300	2,909	text	50%	CC BY-SA 4.0	rajpurkar/squad_v2
cuad	Spans	250	992	text, list	45%	CC BY 4.0	theatticusproject/cuad-qa
cord	Documents	197	1,166	number, list	33%	CC BY 4.0	naver-clova-ix/cord-v2
docs	Documents	400	2,813	text, number, date, time, datetime, duration, list	22%	CC BY 4.0	turion-ai
drop	Numbers	300	1,380	text, number, date, list	0%	CC BY-SA 4.0	ucinlp/drop
tatqa	Numbers	250	1,500	text, number, list	0%	CC BY 4.0	next-tat/TAT-QA
fewnerd	Entities	300	1,517	list	43%	CC BY-SA 4.0	DFKI-SLT/few-nerd

Total: 1,997 requests, 12,277 fields.

Files

• <test>.jsonl.gz: one row per line, {"id", "test", "request", "gold", "meta"}. Engines see only request (state, fields, optional reference_date). gold[field].value is the expected value (null when the document does not determine it), with optional alternatives and, for lists, item_alternatives.
• manifest.json: counts, field types, null share, source dataset and pinned revision, licence and sha256 of every file. Runs record the manifest's sha256, and the board accepts only runs on this exact suite.

The files are rebuilt byte for byte from the public sources with python -m extract_index suite build (seed 20261024, pinned dataset revisions).

Rules

Don't train on these rows or on the docs generator's templates. Use one configuration for every test, with no truncation. See the kit README for the full rules and how to submit.

Licences

Each test keeps its source's licence (table above). docs is our own synthetic set (CC BY 4.0). Please cite the original datasets when you use their tests.