Hugging Face's logo

Datasets:
amazon
/
ConfBench

Tasks:
Document Question Answering
Image-to-Text
Languages:
English
Size:
1K<n<10K
ArXiv:
Tags:
document-understanding
invoice
ocr
noise-augmentation
confidence-calibration
information-extraction
License:
Dataset card Files
xet
 Community
ConfBench / README.md
mmiakashs's picture
mmiakashs
Remove RealKIE-FCC-Verified and RealKIE BibTeX entries from citation section
5909b38 verified
preview code
|
Raw
History Blame Contribute Delete
10.4 kB
metadata 
language:
  - en
license: cc-by-nc-4.0
task_categories:
  - document-question-answering
  - image-to-text
pretty_name: FCC Invoices Verified Augmented
size_categories:
  - 1K<n<10K
annotations_creators:
  - expert-generated
source_datasets:
  - https://indicodatasolutions.github.io/RealKIE/
  - https://huggingface.co/datasets/amazon-agi/RealKIE-FCC-Verified
tags:
  - document-understanding
  - invoice
  - ocr
  - noise-augmentation
  - confidence-calibration
  - information-extraction
  - kie
dataset_info:
  features:
    - name: doc_id
      dtype: string
    - name: pipeline_name
      dtype: string
    - name: original_pdf
      dtype: string
    - name: noisy_pdf
      dtype: string
    - name: ground_truth
      dtype:
        struct:
          - name: document_class
            dtype:
              struct:
                - name: type
                  dtype: string
          - name: inference_result
            dtype:
              struct:
                - name: Agency
                  dtype: string
                - name: Advertiser
                  dtype: string
                - name: GrossTotal
                  dtype: float64
                - name: PaymentTerms
                  dtype: string
                - name: AgencyCommission
                  dtype: float64
                - name: NetAmountDue
                  dtype: float64
                - name: LineItems
                  sequence:
                    struct:
                      - name: LineItemDescription
                        dtype: string
                      - name: LineItemStartDate
                        dtype: string
                      - name: LineItemEndDate
                        dtype: string
                      - name: LineItemDays
                        dtype: string
                      - name: LineItemRate
                        dtype: float64
  splits:
    - name: test
      num_examples: 1346

FCC Invoices Verified Augmented

Dataset Description

FCC Invoices Verified Augmented is a document understanding benchmark dataset consisting of 75 real-world Federal Communications Commission (FCC) invoice documents, each augmented with up to 18 distinct document degradation pipelines. The dataset is designed to support confidence calibration research, OCR robustness evaluation, and key information extraction (KIE) under realistic noise conditions.

Each document has:

  • A clean original PDF
  • Up to 18 noisy versions generated by distinct Augraphy-based degradation pipelines
  • Verified ground-truth entity annotations

Total samples: 1,346 (75 documents × up to 18 noise pipelines)

Quick Start 

from huggingface_hub import snapshot_download

# Download the full dataset locally
local_dir = snapshot_download(repo_id="amazon/ConfBench", repo_type="dataset")

Or clone with git:

git clone https://huggingface.co/datasets/amazon/ConfBench

Or use the provided loader (see load_dataset.py in this release):

from load_dataset import FCCInvoicesDataset

ds = FCCInvoicesDataset(local_dir="/path/to/ConfBench")

for sample in ds:
    print(sample["doc_id"], sample["pipeline_name"])
    print(sample["ground_truth"]["inference_result"]["Agency"])

Dataset Summary

Property Value
Domain Legal / Broadcast Advertising
Document Type FCC Invoice (multi-page PDF)
# Base Documents 75
# Noise Pipelines up to 18 per document
# Total Samples 1,346
Avg Pages per Doc ~2
License CC BY-NC 4.0

Dataset Structure

Repository Layout 

amazon/ConfBench/
└── assets/
    └── {doc_id}/
        ├── original.pdf              # Original clean PDF
        ├── gt.json                   # Ground truth annotations
        ├── metadata.json             # Pipeline manifest
        ├── default/
        │   └── default_noisy.pdf
        ├── archetype3/
        │   └── archetype3_noisy.pdf
        └── ... (16 more pipelines)

Ground Truth Schema (gt.json) 

{
  "document_class": {
    "type": "Invoice"
  },
  "split_document": {
    "page_indices": [0, 1]
  },
  "inference_result": {
    "Agency": "American Media & Advocacy Group",
    "Advertiser": "National Rifle Association",
    "GrossTotal": 15185.0,
    "PaymentTerms": "30 Days",
    "AgencyCommission": 2277.75,
    "NetAmountDue": 12907.25,
    "LineItems": [
      {
        "LineItemDescription": "M-F 1135p-1205a",
        "LineItemStartDate": "10/09/12",
        "LineItemEndDate": "10/09/12",
        "LineItemDays": "-T-----",
        "LineItemRate": 600.0
      }
    ]
  }
}

Metadata Schema (metadata.json) 

{
  "doc_hash": "033f718b16cb597c065930410752c294",
  "original_pdf": "original.pdf",
  "ground_truth": "gt.json",
  "pipelines": [
    {"pipeline_name": "default", "noisy_pdf": "default/default_noisy.pdf"},
    {"pipeline_name": "archetype3", "noisy_pdf": "archetype3/archetype3_noisy.pdf"}
  ]
}

Noise Pipelines

18 distinct Augraphy-based degradation pipelines covering a wide range of real-world scan/print artifacts:

Augraphy Archetypes (pre-built pipelines)

Pipeline Description
default Balanced general-purpose degradation
archetype3 Heavy post-processing effects
archetype4 Minimal geometric distortions
archetype7 Color and lighting variations
archetype9 Texture-based degradations
archetype10 Scanner artifact simulation
archetype11 Complex multi-phase pipeline

Custom Pipelines (research-designed)

Pipeline Key Augmentations Simulates
custom12 DirtyDrum + DirtyRollers Scanner roller artifacts
custom13 Stains + Folding Physical document damage
custom14 BleedThrough + InkMottling Ink bleed and mottling
custom15 Moire + ColorPaper Scanning/aging effects
custom16 ShadowCast + LightingGradient Uneven lighting
custom17 Jpeg + SubtleNoise Compression artifacts
custom18 Geometric + PageBorder Alignment issues
custom19 BindingsAndFasteners + Letterpress Binding shadows
custom20 Brightness + BadPhotoCopy Photocopy quality
custom21 WaterMark + NoisyLines Overlaid artifacts
custom22 Dithering + DotMatrix Dot-matrix printing

Entity Fields

Field Type Description
Agency string Advertising agency name
Advertiser string Client/advertiser name
GrossTotal float Total invoice amount (USD)
PaymentTerms string Payment terms (e.g., "30 Days")
AgencyCommission float Agency commission amount (USD)
NetAmountDue float Net amount after commission (USD)
LineItems list Individual line items
LineItemDescription string Program/slot description
LineItemStartDate string Airing start date (MM/DD/YY)
LineItemEndDate string Airing end date (MM/DD/YY)
LineItemDays string Days of week pattern (e.g., "-T-----")
LineItemRate float Cost per line item (USD)

Usage

Loading with huggingface_hub 

from huggingface_hub import snapshot_download
import json
from pathlib import Path

local_dir = Path(snapshot_download(repo_id="amazon/ConfBench", repo_type="dataset"))
assets_dir = local_dir / "assets"

doc_ids = [p.name for p in assets_dir.iterdir() if p.is_dir()]

# Load ground truth for a single document
def load_gt(doc_id):
    return json.loads((assets_dir / doc_id / "gt.json").read_text())

# List noisy PDFs for a document
def list_pipelines(doc_id):
    meta = json.loads((assets_dir / doc_id / "metadata.json").read_text())
    return meta["pipelines"]

Using the Provided Loading Script 

# See load_dataset.py in this release
from load_dataset import FCCInvoicesDataset

ds = FCCInvoicesDataset(local_dir="/path/to/ConfBench")

# Iterate over all (doc, pipeline) pairs
for sample in ds:
    print(sample["doc_id"], sample["pipeline_name"])
    print(sample["ground_truth"]["inference_result"]["Agency"])

Intended Use

This dataset is intended for:

  1. Confidence calibration research — Measuring model confidence under varying degrees of document degradation.
  2. OCR robustness evaluation — Benchmarking OCR and KIE systems on realistic noisy documents.
  3. Document understanding — Evaluating models on structured information extraction from invoices (evaluation/benchmark use only; no training split is provided).
  4. Noise impact analysis — Studying how specific noise types affect extraction accuracy per field.

Source Data

The 75 clean FCC invoices used in this dataset are taken from RealKIE-FCC-Verified, which re-annotated the FCC Invoices subset of the original RealKIE benchmark (Townsend et al., 2024) to fix two issues in the original annotations:

  1. Line Item Grouping — fields previously treated as independent entries are grouped within each individual line item, aligning annotations with real-world invoice structure.
  2. Annotation Corrections — erroneous values in the original annotations were corrected.

Starting from these 75 verified documents and their corrected gt.json ground truth, this dataset adds noise augmentation using the Augraphy library with OCR-safe parameter settings, producing up to 18 degraded variants per document for confidence calibration and OCR robustness research.

How to Cite This Dataset

If you use this dataset, please cite:

@misc{islam2026confidencecalibration,
  title     = {FCC Invoices Verified Augmented: A Benchmark for Confidence Calibration in Document Understanding under Noise},
  author    = {Md Mofijul Islam and Mohammad Rostami and others},
  year      = {2026},
  note      = {Dataset available at Hugging Face},
  url       = {https://huggingface.co/datasets/amazon/ConfBench}
}

License

This dataset is released under CC BY-NC 4.0. The original FCC invoice documents are public records.