README.md

metadata

license: cc-by-sa-4.0
language:
  - en
task_categories:
  - token-classification
  - question-answering
pretty_name: typed-extraction
size_categories:
  - 10K<n<100K
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train.parquet
      - split: test
        path: data/test.parquet

typed-extraction

A benchmark for typed extraction over a document: one piece of text in, a named map of typed fields asked at once, and for each field either a typed value or null when the document does not say.

It follows the System One extraction request shape (state, fields, reference_date; answers with value, span, p_null, confidence), so a case can be replayed against any API that implements it. It is the sibling of typed-decisions. It is an independent benchmark. It is not affiliated with any vendor of extraction models.

Leaderboard

test split: 1,690 cases, 17,733 fields. This table lists models scored zero-shot: they never saw train. Models fitted or trained on train are in the second table below; the two tables are not comparable.

Headline is exact match averaged over the 15 families; fields with a value is the same over fields whose gold is not null; computed is exact match on the 1,926 fields whose answer has to be worked out (totals, counts, day spans). Every model sees the same request: the document and the fields with their instructions, nothing else.

#	Model	Kind	Headline ↑	Fields with a value ↑	Hallucination ↓	Miss ↓	Computed ↑	p_null AUROC ↑
1	ScaleDown Extract	extraction API	0.782	0.723	0.117	0.200	0.570	–
2	UltraBERT-150M clairic-extraction	span extractor, open 150M encoder	0.616	0.548	0.141	0.297	0.013	0.815
3	Schematron V2 Small	extraction LLM	0.562	0.529	0.445	0.303	0.134	–
–	Always null (answers nothing)	reference	0.199	0.000	0.000	1.000	0.000	–
–	Regex rules	reference	0.129	0.152	0.993	0.000	0.026	0.503

Fitted or fine-tuned on train

These models were trained on this benchmark's train split, so they are not zero-shot and their scores are not comparable with the table above. The three test-only families never appear in train, so they still test transfer.

#	Model	Kind	Headline ↑	Fields with a value ↑	Hallucination ↓	Miss ↓	Computed ↑	p_null AUROC ↑
1	adaptive-classifier + MiniLM-L6	candidate classifier, trained on train	0.315	0.166	0.109	0.600	0.005	0.789
2	adaptive-classifier + bge-small	candidate classifier, trained on train	0.291	0.143	0.103	0.600	0.008	0.788
–	Always null (answers nothing)	reference	0.199	0.000	0.000	1.000	0.000	–
–	Regex rules	reference	0.129	0.152	0.993	0.000	0.026	0.503

Notes on the rows

• ScaleDown Extract is the bar. It returns values that are not in the text (it does the arithmetic), and its weak spots are ExtractBench's long nested schemas and relative dates.
• The adaptive-classifier baselines pick one of 40 candidate spans per field with a single yes/no classifier, so they rarely hallucinate but cannot compute, and for fields that have a value the right answer is among their candidates only 49% of the time. That ceiling, not the classifier, is what holds them back. They are published as reproducible open baselines, not as competitive extractors.
• UltraBERT-150M clairic-extraction (submitted by sileod, discussion #1) is the first open-weight row. It tags answer spans per field with no LLM and maps them to types in plain code, so it rarely computes values (0.013). Its confidence is the raw span probability (ECE 0.42), but it ranks nulls well. We re-scored the published predictions with the unmodified score.py and got the same numbers. Its training data includes the SQuAD 2.0, DROP and CUAD train splits; 23 test documents share text with it, and dropping them leaves the headline at 0.616.
• Always null answers nothing; it scores 0.199 because 31% of fields are absent. A model below it is doing more harm than good.
• Remote models were called once, at temperature 0, with responses cached. LLMs got a JSON schema with one nullable property per field; documents longer than their context were cut (5 cases for Schematron), and that is counted.
• Rows for LFM2.5 and Clairic will be added when their runs finish.

Scoring your model

pip install numpy pyarrow huggingface_hub
huggingface-cli download LocalLLaMA/typed-extraction --repo-type dataset --local-dir typed-extraction
python typed-extraction/score.py predictions.jsonl --by-family

predictions.jsonl has one line per test case, {"id": ..., "answers": {"<field>": {"value": ..., "p_null": ..., "confidence": ...}}}, with null (or [] for lists) for fields the document does not state. p_null and confidence are optional and enable the calibration metrics. Please report the headline, hallucination and miss together; any one alone can be gamed. To be added to the leaderboard, open a discussion on this dataset with the scorer's output and a link to the model, and say whether it was zero-shot or trained on train.

What is different about it

• The gold is fact, not a teacher's opinion. Every answer is a human annotation from an open dataset, the record a constructed document was rendered from, or a Wikidata fact verified against the article text. No model labelled anything, so there is no teacher ceiling and frontier models can be ranked.
• Absent fields are first-class. 31% of test fields have null gold, and the headline reports the hallucination rate (a value given for an absent field) separately from the miss rate.
• Typed answers. Fields are text, number, date, time, datetime, duration or list. A value has to be the right type, not a string that happens to contain it. Numbers may be computed (totals, counts, a deposit stated as "two months' rent"), and dates may be relative ("delivered last Tuesday", resolved against reference_date).
• Held-out families. Three constructed families (real-estate listings, meeting notes, medical appointment letters) appear only in the test split, so a score on them cannot come from memorised templates.
• Variation. Constructed cases draw field names and instructions from synonyms. They shuffle field order, ask for a plausible field the document does not contain, and pad 30% of documents to 3–12k characters with decoy values of the same shape.

Splits

	test	train
cases	1,690	9,642
fields	17,733	74,014
null gold	31.1%	19.7%
held-out family fields	2,380	0

Test fields by family:

family	origin	cases	fields	what it tests
squad2	converted	200	1,908	spans; unanswerable questions as null
drop	converted	198	1,684	numbers, mostly computed
cuad	converted	102	4,182	41 clause lists per long contract (median ~26k chars)
extractbench	converted	30	2,115	nested schemas on long PDFs (10-Q, credit agreements, papers…), test only
wikipedia	distant	300	964	real articles, Wikidata values verified in the text
claim_email	constructed	80	617	relative dates, computed totals and day counts
invoice	constructed	80	681	due dates from payment terms, totals, line items
lease	constructed	80	770	end dates, deposits stated as multiples of rent
itinerary	constructed	80	608	datetimes, flight lists, journey minutes
shipping_notice	constructed	80	607	business-day ETAs, unit conversion
job_posting	constructed	80	609	salary ranges ("$85k–$110k"), skill lists
bank_statement	constructed	80	608	closing balances, largest debit, counts
real_estate_listing	constructed, held out	100	961	price per square foot, open-house datetimes
meeting_notes	constructed, held out	100	659	durations, "same time next week"
medical_appointment	constructed, held out	100	760	arrive-by times, absent copays and referrals

Test fields by type:

type	fields
list	6,854
text	5,278
number	3,480
date	1,119
duration	442
datetime	380
time	180

Present test gold is 8,192 verbatim spans, 2,020 normalised (written in another form), 1,926 computed and 86 relative.

Scoring

typed_extraction.metrics (numpy only) scores each field exact (0/1) and partially:

• text: SQuAD-normalised exact match; partial credit is token F1.
• number / duration: relative tolerance 1e-3, or rounding at the field's decimals.
• date / time / datetime: equal ISO values.
• list: multiset F1. A text item matches at token Jaccard ≥ 0.5 (CUAD's convention), and exact means F1 = 1.
• null: right only when the prediction is null (an empty list for lists).

Headline: the mean over families, since families differ about 10× in field count. It is reported with hallucination rate, miss rate, p_null AUROC, ECE, and accuracy at 90% coverage (answering only the most confident 90%) for predictors that report confidence.

Gold quality

200 test fields were hand-checked against their documents, weighted towards null and derived gold. 2 were wrong (1.0%, 95% CI 0.3–3.6%). Both were fixed, together with every recurring issue the audit found:

• CUAD exact clause boundaries, by moving to Jaccard list matching;
• honorific-only differences;
• ExtractBench instructions that asked for 'NA' where the gold is null;
• ambiguous slash dates;
• two malformed ExtractBench fields.

Null gold had no errors in the sample.

Known limitations:

• The Wikipedia slice is thin: mostly 3 fields per case, and its few nulls are all death dates.
• ExtractBench text comes from pdftotext. 5 of its 35 PDFs (image-only resumes) are excluded, and its LLM-judged metrics are not reproduced, so its scores are not comparable to the ExtractBench paper.
• SQuAD 2.0 and DROP are in most models' pretraining data.
• Itinerary durations use local clock times; the instructions say so.

Sources and licences

source	licence
SQuAD 2.0	CC BY-SA 4.0
DROP	CC BY-SA 4.0
CUAD v1	CC BY 4.0
ExtractBench, commit ca81763	MIT
Wikipedia text	CC BY-SA 4.0
Wikidata	CC0
constructed families	generated for this dataset

The dataset as a whole is released under CC BY-SA 4.0, as the share-alike sources require.

Format

Each row is one case:

column	content
id, family, origin, held_out	identifiers
state	the document
reference_date	the date relative expressions resolve against, if any
fields	JSON: name → {type, instructions, options}
gold	JSON: name → {value, source, alts, spans}

spans are [start, end) character offsets into state.