Dataset Viewer
Auto-converted to Parquet Duplicate
doc_id
stringlengths
7
69
category
stringclasses
10 values
source
stringclasses
2 values
original_format
stringclasses
5 values
schema
stringclasses
10 values
document
stringlengths
175
4M
expected
stringlengths
113
2.58k
edgar_contract_001
contracts
real
HTML
contract.yaml
Exhibit 10.24(c) MASTER ASSIGNMENT, ASSUMPTION AND ACKNOWLEDGEMENT AGREEMENT This MASTER ASSIGNMENT, ASSUMPTION AND ACKNOWLEDGEMENT AGREEMENT (this “Agreement”), dated as of the 30th day of June, 2008, by and between Equitable Resources, Inc., a Pennsylvania corporation formed in 1926 (“Assignor”) and Equitable Resou...
{"contract_type": "Other", "effective_date": "2008-06-30", "governing_law": "Pennsylvania", "parties": ["Equitable Resources, Inc. (organized in 1926)", "Equitable Resources, Inc. (organized in 2008)"], "termination_date": null}
edgar_contract_002
contracts
real
HTML
contract.yaml
Exhibit 10.106 UPON RECORDING RETURN TO: Williams, Rentz & Moulton, P.C. 412 Ocean Boulevard St. Simons Island, Georgia 31522 DEED TO SECURE DEBT AND SECURITY AGREEMENT THIS DEED TO SECURE DEBT AND SECURITY AGREEMENT (hereinafter referred to as this “Deed”) made and entered into as of the 7th day of October, ...
{"contract_type": "Credit Agreement", "effective_date": "2022-10-07", "governing_law": "Georgia", "parties": ["MHP PURSUITS LLC", "RICHARD SMITH", "ANNETTE SMITH"], "termination_date": "2042-11-01"}
edgar_contract_003
contracts
real
HTML
contract.yaml
Exhibit 10.2 EXECUTIVE EMPLOYMENT AGREEMENT This Executive Employment Agreement (the “Agreement”) is made and entered into as of July 10, 2017 by and between James E. Mead (the “Executive”) and Alexander & Baldwin, Inc. (the “Company”) to become effective upon the Executive’s commencement of employment with the C...
{"contract_type": "Employment Agreement", "effective_date": "2017-07-10", "governing_law": "Hawaii", "parties": ["James E. Mead", "Alexander & Baldwin, Inc."], "termination_date": "2018-07-10"}
edgar_contract_004
contracts
real
HTML
contract.yaml
"EXHIBIT (10)(IV)\n\nOUTPARCEL GROUND LEASE\n\nbetween\n\nJPC MONROE, LLC, Landlord\n\nand\n\nBANK O(...TRUNCATED)
"{\"contract_type\": \"Lease\", \"effective_date\": \"2005-09-13\", \"governing_law\": \"North Carol(...TRUNCATED)
edgar_contract_005
contracts
real
HTML
contract.yaml
"Exhibit 10.56.3\n\n2004 Genworth Financial, Inc.\n\nOmnibus Incentive Plan\n\nStock Option Award Ag(...TRUNCATED)
"{\"contract_type\": \"Other\", \"effective_date\": null, \"governing_law\": \"Delaware\", \"parties(...TRUNCATED)
edgar_contract_006
contracts
real
HTML
contract.yaml
"|   | X=                  Y(A-B)_ |\n|--------------|------------(...TRUNCATED)
"{\"contract_type\": \"Other\", \"effective_date\": null, \"governing_law\": null, \"parties\": [\"D(...TRUNCATED)
edgar_contract_007
contracts
real
HTML
contract.yaml
"Exhibit 10.3\n\nREGISTRATION RIGHTS AGREEMENT\n\nTHIS REGISTRATION\nRIGHTS AGREEMENT (this “Agree(...TRUNCATED)
"{\"contract_type\": \"Other\", \"effective_date\": \"2022-02-10\", \"governing_law\": \"New York\",(...TRUNCATED)
edgar_contract_008
contracts
real
HTML
contract.yaml
"EXHIBIT 10-3\n\nGUARANTY\n\nIN CONSIDERATION of the sum of One Dollar ($1.00), cash in hand paid, a(...TRUNCATED)
"{\"contract_type\": \"Other\", \"effective_date\": \"2006-07-13\", \"governing_law\": \"Florida\", (...TRUNCATED)
edgar_contract_009
contracts
real
HTML
contract.yaml
"Exhibit 10.1\n\nCOMMON STOCK PURCHASE AGREEMENT\n\nAGREEMENT entered into as of the 19th day of Au(...TRUNCATED)
"{\"contract_type\": \"Other\", \"effective_date\": \"2015-08-19\", \"governing_law\": \"New York\",(...TRUNCATED)
edgar_contract_010
contracts
real
HTML
contract.yaml
"Exhibit 10-1\n\nPREPARED BY AND RETURN TO:\n\nMITCHELL S. GOLDMAN\n\n96\nWILLARD STREET, SUITE 302\(...TRUNCATED)
"{\"contract_type\": \"Other\", \"effective_date\": \"2010-07-30\", \"governing_law\": \"Florida\", (...TRUNCATED)
End of preview. Expand in Data Studio

FieldBench Corpus

A cross-domain, field-level benchmark for schema-driven document extraction (document → structured JSON). 1,441 documents across 10 categories with per-field ground truth, released so extraction-accuracy claims become falsifiable and comparable.

Load

from datasets import load_dataset
ds = load_dataset("fieldbench/corpus")["test"]
ex = ds[0]
# ex["document"]  -> the markdown representation to extract from
# ex["expected"]  -> ground-truth field map (JSON string; json.loads it)
# ex["category"], ex["source"] ("real"|"synthetic"), ex["schema"]

Score predictions with the canonical scorer:

pip install fieldbench
fieldbench score --corpus <corpus-checkout> --results <your-predictions>/

Fields

field description
doc_id document identifier
category one of 10 categories (sec_filings, invoices, medical_records, …)
source real or syntheticalways report results stratified by this
original_format how the document reached the extractor (see composition note)
document the markdown representation to extract from
expected ground-truth {field: value} map, serialized as a JSON string
schema schema file name defining the fields for this category

Important caveats

  • ~90% of documents are extraction-from-clean-text, not rendered-page extraction (only 9.8% came from an image/PDF). Parse-stage difficulty is largely absent by construction — scope accuracy claims accordingly. See the composition table in the corpus repo.
  • Synthetic documents (~48%) overestimate accuracy relative to real ones. Never report a synthetic-inclusive number without the real/synthetic split.
  • Licensing is per-source (see ATTRIBUTION.md): synthetic CC0; SEC EDGAR public disclosure; SROIE CC BY 4.0; MTSamples educational-use with attribution; Caselaw Access Project public domain; government forms public domain. ACORD/ISO copyrighted forms are not included (those categories use synthetic equivalents).

Citation

See CITATION.cff in the corpus repo.

Downloads last month
68