sujimart/testing-kie-1

ConfBench

A benchmark dataset for evaluating noise robustness in Intelligent Document Processing (IDP) pipelines. Each document is represented as a clean original plus multiple degraded variants produced by different noise pipelines, all sharing the same ground truth.

Designed for ingestion by the AWS IDP Accelerator Test Studio.

Dataset Structure

Each row is one (document × noise_variant) pair. Parquet metadata lives in data/, PDFs in pdfs/.

Column	Type	Description
id	string	{doc_hash}__{noise_variant}.pdf — unique row key, matches filename in pdfs/
noise_variant	string	original, default, archetype3, custom12, …
page_count	int	Number of pages in the document
json_response	struct	Ground truth as a parsed dict, consistent with RealKIE-FCC-Verified

Ground Truth Format

json_response contains the extracted invoice fields, consistent with RealKIE-FCC-Verified:

{
  "Agency": "...",
  "Advertiser": "...",
  "GrossTotal": 0.0,
  "PaymentTerms": "...",
  "AgencyCommission": 0.0,
  "NetAmountDue": 0.0,
  "LineItems": [
    {
      "LineItemDescription": "...",
      "LineItemStartDate": "...",
      "LineItemEndDate": "...",
      "LineItemDays": "...",
      "LineItemRate": 0.0
    }
  ]
}

Noise Variants

Each document is processed through multiple degradation pipelines:

• original — clean source PDF
• default — baseline noise pipeline
• archetype* — archetype-based noise profiles (archetype3, archetype4, archetype7, archetype9, archetype10, archetype11)
• custom* — custom noise configurations (custom12–custom23)

Not every document has every variant — the set may differ across documents.

Usage

from datasets import load_dataset

ds = load_dataset("sujimart/testing-kie-1")
print(ds["test"][0])

# All variants for a specific document
doc_hash = "033f718b16cb597c065930410752c294"
variants = [r for r in ds["test"] if r["id"].startswith(doc_hash)]

# Access ground truth (json_response is a dict, not a string)
gt = variants[0]["json_response"]
print(gt["Agency"])

To download a PDF directly:

from huggingface_hub import hf_hub_download

pdf_path = hf_hub_download(
    repo_id="sujimart/testing-kie-1",
    filename=f"pdfs/{row['id']}",
    repo_type="dataset",
)