FairForm-Bench

A benchmark for measuring document extraction accuracy under demographically-relevant document quality variation, built for the FairForm research project (AIF 2026 Capstone).

Generated: 2026-09-28

---

Dataset Summary

FairForm-Bench contains 5,000 synthetic loan application form images (1,000 unique form records × 5 controlled degradation levels), rendered from realistic template layouts and systematically degraded to simulate real-world document scan quality disparities.

The dataset supports the FairForm research question: do systematic extraction errors — driven by document quality disparities correlated with applicant demographics — amplify bias in downstream automated credit decisions? See the FairForm project for the full pipeline that uses this benchmark (Agent 2 extraction experiments, Experiments 1–3).

Total size on disk: ~30,492.3 MB (images + ground truth JSON)

---

Methodology

Generation Pipeline

1. Record sampling — 1,000 unique loan application records were generated using Faker for realistic applicant names, addresses, and employer names, combined with randomly sampled financial fields (income, loan amount, DTI) within realistic ranges.
2. Form rendering — Each record was rendered as a filled PDF using one of 3 ReportLab form templates (see Template Distribution below), then converted to a 300 DPI PNG via pdf2image.
3. Degradation injection — Each rendered form was processed through 5 controlled degradation levels using OpenCV, producing 5 image variants per unique form (see Degradation Levels below).
4. Ground truth capture — The exact field values used to populate each form are stored as a paired JSON file, enabling exact-match extraction accuracy evaluation.

Degradation Levels

Level	Count	Description
clean	1,000	Original 300 DPI PDF render. No modifications.
light	1,000	Mild JPEG compression (quality=60) + subtle Gaussian blur.
medium	1,000	Random rotation ±2° + additive Gaussian noise (σ=12).
heavy	1,000	3-generation photocopier simulation: iterative JPEG + downsample.
extreme	1,000	Heavy noise + contrast wash + horizontal fax streaks + hard JPEG.

Degradation severity was calibrated by visual inspection against the FUNSD benchmark's real noisy scanned documents — Levels 1–2 are intended to approximate FUNSD's ~100 DPI scan conditions.

Template Distribution

Template	Unique Forms	% of Dataset
Heritage Federal Bank (boxed-field, formal)	338	33.8%
LoanPro Solutions (underline-field, modern)	334	33.4%
Regional Loan Authority (dense, government style)	328	32.8%

Financial Field Ranges

Field	Min	Max
Annual income	$28,021	$279,981
Loan amount requested	$60,543	$900,000
Debt-to-income ratio	15.0%	54.9%

---

Dataset Structure

fairform-bench/
├── images/
│   ├── form_0001_L0.png      # clean
│   ├── form_0001_L1.png      # light degradation
│   ├── form_0001_L2.png      # medium degradation
│   ├── form_0001_L3.png      # heavy degradation
│   └── form_0001_L4.png      # extreme degradation
│   ...
├── ground_truth/
│   ├── form_0001.json        # one file per unique form (shared across all 5 levels)
│   ...
├── metadata.csv               # full dataset index
└── README.md                  # this file

Ground Truth Schema

Each ground_truth/form_XXXX.json contains:

Field	Type	Description
applicant_name	string	Faker-generated name
date_of_birth	string	MM/DD/YYYY
annual_income	int	USD
employment_status	string	One of 5 categories
employer_name	string	Faker-generated company, or "Self-employed"
loan_amount_requested	int	USD
property_address	string	Faker-generated US address
loan_purpose	string	One of 5 categories
debt_to_income_ratio	float	Decimal, e.g. 0.38
collateral_type	string	One of 5 categories
co_applicant_name	string | null	Present in ~30% of records
signature_date	string	MM/DD/YYYY
template_id	int	1, 2, or 3
demographic_group	string	See Demographic Labels below

metadata.csv Columns

form_id, level, level_name, image_path, gt_path, template_id, applicant_name, annual_income, loan_amount_requested, loan_purpose, employment_status, debt_to_income_ratio, demographic_group

---

Demographic Labels

demographic_group value	Unique Forms
unassigned	1,000

Important: demographic_group is set to "unassigned" in this release of the dataset. In the full FairForm pipeline, this field is intended to be overridden with applicant_race_1 values from linked HMDA source records — demographic labels are never inferred from applicant names, addresses, or any visual signal in the form image itself. Users who wish to run demographic-stratified experiments must perform this join against HMDA data themselves; see the FairForm project repository for the join procedure.

---

Known Limitations

• Names are English-locale only. The generator was originally designed to vary Faker locales (e.g., es_MX, zh_TW, pt_BR, ar_AA) to produce visual name diversity across the dataset. In this generated release, only the en_US locale was used — all applicant names are Anglophone. This does not by itself limit the dataset's use for extraction-accuracy or degradation-level research (which does not depend on name script), but it does mean the dataset does not currently exercise VLM/OCR performance on non-Latin scripts or transliterated names. If your use case requires script diversity, regenerate the dataset with NAME_LOCALES restored to its multi-locale configuration before use.
• No real loan form images. All forms are synthetically generated. No public dataset of real loan application images with demographic labels exists, which is the underlying reason a synthetic approach was used. Results on this benchmark should be validated against real-world data where possible before drawing operational conclusions.
• Only 3 form templates. Real-world institutional loan forms exhibit far more layout diversity than the 3 templates used here.
• Degradation is synthetic, not sampled from real scans. The 5 degradation levels are OpenCV-simulated (blur, noise, compression, rotation, fax-artifact streaks) rather than derived from a distribution of real scanned document quality. Calibration against FUNSD provides partial external validation only.
• Demographic correlation is not baked into the images. This release does not itself encode any correlation between demographic group and degradation level — that correlation is applied at experiment time by the FairForm pipeline (see Experiment 3 in the main project), not within this dataset release.

---

Licensing

Released under CC-BY-4.0. You are free to share and adapt this dataset for any purpose, including commercially, provided appropriate credit is given.

Citation

If you use FairForm-Bench, please cite the FairForm project:

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

Contact

For questions about this dataset or the FairForm project, please open an issue on the project repository.