Research datasets and data points

AMLOK’s public-record collection preserves each source release and the meaning of its fields. The tables below describe acquired records, not estimates of records available elsewhere. Download the collection summary, source-table counts and acquisition manifest.

Collected datasets

Dataset and scope	Observation count	Data points available	Interpretation
SBA 7(a) and 504, six complete releases spanning FY1991 onward; snapshot September 30, 2026	2,207,426 loan-program rows	Approval date, first disbursement, approval amount, guarantee amount, program, industry, business age, initial rate, rate type, reported status and selected outcome dates; 42 fields for 7(a), 39 for 504.	1,897,410 rows contain first-disbursement dates. Approvals, cancellations, funded loans and resolved loans remain distinct.
Investissement Québec, 2023-24 intervention disclosure	1,017 rows	Company text, assistance type, financing amount, PDF page and row location.	993 amounts recovered through OCR. Five anchors visually checked; other recovered amounts require review.
EDC captured disclosure page	30 rows	Signing date, country, counterparty, purpose, industry, sector of interest, cleantech indicator, product, description, CAD amount band and Canadian company.	One visible table page, not the entire archive. Published bands remain bands.
Kinross and Lundin issuer disclosures	2 announcements	Parties, facility type, principal or capacity, currency, term, maturity, conditional stages, rate basis and use of proceeds where disclosed.	A term-loan announcement and a facility amendment; neither supplies a complete repayment ledger.
HMDA 2024, New York, action taken = 1	208,818 mortgage originations	99 source fields, including amount, purpose, loan type, income, debt-to-income band, combined loan-to-value, property value, rate, costs and reporting institution.	Privacy-modified public values; geographic cohort only. No subsequent repayment outcome supplied.
SEC financial-statement data, 2026 Q3	3,905,397 numeric facts from 6,516 filings	Accession, issuer CIK, tag, taxonomy version, period, quarters, unit, value, segment and co-registrant context.	The archive also has 87,871 tag definitions and 731,170 presentation rows. Facts are not loans or unique businesses.

The 1,049 Canadian-source disclosure rows comprise the 1,017 interventions, 30 EDC disclosures and two issuer announcements. Do not sum all datasets into a unique-deal count. A single issuer, facility or financial period can appear in several records.

Source-level fields and canonical attributes

The source-field dictionary contains 424 source/field definitions, including repeated fields tracked separately by release. A source field retains its original name, definition, populated-row count and denominator. The canonical dictionary groups the concepts used for labeling:

Attribute family	Required context	Labeling rule
Identity	Source identifier, named party, institution or issuer, jurisdiction and role	Similar names generate candidates; they do not prove legal identity or ownership.
Financing value	Currency, amount or band, amount basis and source date	Separate approval, commitment, availability, drawdown and outstanding balance.
Time	Approval, signing, first disbursement, publication, reporting period and acquisition	Preserve distinct dates and source precision. Do not backdate availability.
Instrument	Program, loan/facility type, assistance type and conditional terms	Preserve source-specific meanings before mapping to a shared category.
Outcome	Source status, event date, horizon and observation eligibility	A charge-off label is not a universal default label; unknown status remains unknown.
Financial fact	Tag, taxonomy, unit, period, quarters, segment and filing	Do not combine facts with incompatible periods or overlapping consolidation scope.
Provenance	Original URL, file hash, row/page locator and extraction method	Every derived value must resolve to the preserved source version.

Review and labeling

The internal review workbench contains 4,149 records: 3,149 loan/disclosure observations, 500 mortgage records and 500 financial facts. It is a review subset. The complete source files remain in the controlled research workspace. Each record has source values, a suggested label, review status, two independent reviewer fields, a final label and an adjudicator. Suggested labels do not become verified labels until review is complete. Missing values stay missing. The SBA review preview uses a deterministic hash rule capped at 350 rows per release and is not a representative population sample.

Task eligibility

SBA records support a scoped retrospective charge-off study. HMDA supports origination-field analysis, not repayment prediction without linked outcomes. SEC records support financial-fact interpretation and document reasoning. Canadian disclosures support extraction, financing semantics and entity/facility linkage. These releases do not contain linked private bank ledgers, payment-processor events or complete ecommerce/accounting histories. Transaction-understanding and longitudinal monitoring evaluations require separately authorized evidence with suitable labels. Preserve these task boundaries when constructing training cohorts.

Original publishers

Source routes: SBA, Investissement Québec, EDC, HMDA and SEC. The manifest records exact downloaded URLs and file hashes. Source-specific access, attribution and use conditions remain attached to each release.

Research program

Data and models

Risk analysis

Evaluation and evidence

Research datasets and data points

Collected datasets

Source-level fields and canonical attributes

Review and labeling

Task eligibility

Original publishers

​Collected datasets

​Source-level fields and canonical attributes

​Review and labeling

​Task eligibility

​Original publishers

Collected datasets

Source-level fields and canonical attributes

Review and labeling

Task eligibility

Original publishers