FairForm-Bench

Synthetic loan application forms built from real HMDA 2023 applicants, rendered at five controlled degradation levels, for measuring whether document-extraction errors amplify demographic bias in automated credit decisions.

Generator version 2.0.0 · generated 2026-10-24 · 1,000 forms × 5 levels = 5,000 8-bit grayscale PNG images (~13 GB) · project repository

---

Summary

Each form is a 300 DPI US-Letter loan application whose financial fields come from a real applicant in the HMDA 2023 public loan data. Every form is rendered once and then degraded at five levels, so the same form can be compared from clean (L0) to extreme (L4). The ground truth gives, for every field, the canonical value, the exact text printed and its bounding box at every level, plus the applicant's race label and the actual lending decision.

The dataset supports the FairForm research question: do extraction errors, driven by document-quality differences that correlate with demographics, amplify bias in automated credit decisions? That's measured with the Extraction Bias Gap (EBG): the change in a group's denial rate when a credit model scores extracted values instead of the true ones.

Every form carries a real HMDA applicant's race label (demographic_group). Labels come only from the applicant's own HMDA field (applicant_race_1), never from names, addresses or anything visual.

How the forms were made

1. Applicants: 200 per race, drawn (race-stratified, seeded) from the test split of the FairForm credit model's HMDA sample, so no form shows a record the credit model was trained on. Income, loan amount, property value and debt-to-income ratio are the applicant's own (DTI is reported by HMDA as a bracket; the form shows a value inside it). Loan term and the conforming-loan flag are not printed but are in the ground truth.
2. Other fields: name, date of birth, address, employer, purpose and collateral are synthetic (Faker, English locales only).
3. Variations (seeded per form, recorded in the ground truth and in metadata.csv, and statistically independent of race):

Variation	Forms	What it does
hmda_link	100.0%	Financials (income, loan amount, property value, DTI) and the race label come from a real HMDA 2023 applicant.
consistency	40.8%	Employer rewritten to agree with employment status (e.g. Retired → "Retired", Self-employed → own business).
name_format	47.4%	English name format: middle name/initial, hyphenated or apostrophe surname, suffix, or long name (en_US / en_GB / en_IE, ASCII).
missing	21.4%	Fields left off the form: blank, "N/A", "—" or "see attached".
format_variation	44.5%	Dates, amounts and DTI printed in a lossless alternative format (e.g. Jan 15 1985, $85k, 0.384).

4. Rendering: one of three ReportLab templates, rasterised at 300 DPI.
5. Degradation: five levels, applied to the same render:

Level	Name	Images	Degradation
L0	clean	1,000	Original 300 DPI render, no changes.
L1	light	1,000	Mild JPEG compression (quality 60) + slight Gaussian blur.
L2	medium	1,000	Random rotation ±2° + Gaussian noise (σ = 12).
L3	heavy	1,000	3-generation photocopy: repeated JPEG + downsample/upsample.
L4	extreme	1,000	Heavy noise, washed-out contrast, fax streaks, hard JPEG.

Applicants per race

demographic_group	Forms	Actual denial rate	Mean population_weight
American Indian or Alaska Native	200	10.5%	1.0
Asian	200	7.0%	8.0
Black or African American	200	12.5%	4.5
Native Hawaiian or Other Pacific Islander	200	9.5%	1.0
White	200	5.5%	51.4

Groups are balanced by design, not by population, so small groups can be measured. Each form's population_weight (the HMDA sample weight) says how many eligible applicants it represents. Weight by it to estimate population-level rates.

Templates

Template	Forms	Share
Heritage Federal Bank (boxed fields, formal)	338	33.8%
LoanPro Solutions (underlined fields, modern)	334	33.4%
Regional Loan Authority (dense, government style)	328	32.8%

Values

Field	Min	Max
Annual income	$2,000	$5,000,000
Loan amount	$5,000	$3,255,000
Property value	$15,000	$5,755,000
Debt-to-income ratio	5.1%	74.7%

Name formats: middle name or initial 33.4%, hyphen 6.9%, apostrophe 3.4%, suffix 6.4% of applicants.

---

Files

fairform-bench/
├── images/form_0001_L0.png … form_0001_L4.png   # one image per form and level
├── ground_truth/form_0001.json                  # one file per form (all levels)
├── metadata.csv                                 # one row per image
└── qa/                                          # contact sheets (one per template)

Ground truth (ground_truth/form_XXXX.json)

Key	Contents
fields	Per field: value (canonical: ISO date, integer USD, decimal DTI), rendered_text (exactly what is printed, captured while drawing) and present (whether it's on the form). Score extraction against rendered_text, normalisation against value.
bboxes	Per level ("0"–"4"), per printed field: [x0, y0, x1, y1] in pixels, origin top-left, following the L2 rotation. Blank fields have no box.
hmda	hmda_row_id, dti_bin, loan_term, conforming_loan_limit, denied (the actual decision), population_weight.
demographic_group	HMDA applicant_race_1 (5 groups).
missing_required	Credit-model inputs left off the form (usually empty).
augmentations	The variations applied to this form, with their parameters.
template_id, page, image_mode, generator_version	Template, page size and DPI, PNG mode, generator release.
flat keys	applicant_name, annual_income, … The canonical values, kept for compatibility.

metadata.csv has one row per image: ids, paths, level, template, the main field values, demographic_group, population_weight, missing_required and one aug_<name> 0/1 column per variation.

---

Using it for bias measurement

• Exclude missing_required forms (19 here) from decision-based metrics. They leave a credit-model input off the form, so they can't be scored and should go to human review instead; they exist to test that path.
• Compare levels on the same form. Degradation is applied identically regardless of race; any demographic correlation with document quality is introduced at experiment time, not baked into the data.
• Control for template. Degradation affects the templates slightly differently (up to ~1.7× pixel change at L1), and the template is recorded per form.
• Re-weight with population_weight for population estimates.

Quality checks

The release was audited with python -m form_generation.audit --release. The checks include:

• every printed value parses back to its canonical value (nothing truncated);
• every printed field's box contains ink at L0;
• the employment fields are consistent;
• the HMDA feature round trip reproduces every form's credit-model inputs;
• the template and every variation are statistically independent of race.

Limitations

• Synthetic forms and synthetic degradation. No public dataset of real loan application images with demographic labels exists, which is why this one is synthetic. Validate conclusions on real documents where possible.
• Three templates, US Letter, typed text. No handwriting, signatures, checkboxes, stamps or phone-camera capture yet.
• English only. Names come from en_US / en_GB / en_IE locales, in ASCII; non-Latin scripts and multilingual forms are not covered.
• HMDA scope. Conventional, owner-occupied, first-lien home-purchase applications from 2023; DTI is printed from HMDA's reported bracket.

Changelog

• 2.0.0: forms carry real HMDA 2023 applicants (race-stratified, with population weights, property value printed); the per-field value / rendered_text / present ground truth; bounding boxes; the consistency, name_format, missing and format_variation variations; grayscale images; a release audit. Replaces the partial 1.x upload, which is kept at the Hub tag v1.0-partial.
• 1.1: ground truth matches the image exactly; fully seeded generation.
• 1.0: initial release (synthetic financials, no HMDA link).

License and citation

Released under CC-BY-4.0. HMDA data is public-domain US government data.

@misc{fairform2026,
  title  = {FairForm: A Vision-Based Multi-Agent Audit Framework for Quantifying
             Extraction-Induced Demographic Bias in Automated Loan Processing},
  author = {Bikal Poudel},
  year   = {2026},
  note   = {AIF 2026 Capstone},
  url    = {https://huggingface.co/datasets/bikalpoudel/fairform-bench}
}

Questions and issues: project repository.