DevLCodes/three-way-matching-dataset

Three-Way Matching Multimodal Agent Benchmark (2,500 Cases / 7,500 Documents)

A multimodal Accounts Payable (AP) & Procurement Three-Way Matching benchmark dataset designed for evaluating LLM/VLM agents, OCR pipelines, and document-understanding workflows across Purchase Orders (PO), Warehouse Receipts / Goods Receipt Notes (GRN), and Vendor Invoices.

Dataset Highlights

• 2,500 Complete Procurement Cases (train: 2,000 cases, test: 500 cases — stratified across all 13 matching & exception scenarios)
• 7,500 Multimodal Documents (train: 6,000 docs, test: 1,500 docs — 1,230.7 MB total)
• Mixed Formats Inside Every Triplet: Every standard 3-document packet combines Word (.docx), PDF (.pdf), and Scanned Image (.jpg / .png), rotating which document role receives which format across cases.
• Both Vector & Scanned PDFs: Includes selectable digital ERP PDFs (vector_text_pdf) as well as rasterized scanned PDFs (scanned_image_pdf).
• Realistic Visual & Dock Artifacts: Image and scanned-PDF documents feature page skew (±0.5° to ±2.1°), scanner drop shadows, rubber ink stamps (RECEIVED & VERIFIED, SHORT SHIPMENT NOTED, DAMAGED GOODS REJECTED, AP RECEIVED), vector barcodes, paper speckles/creases, and ballpoint handwritten dock clerk notes + cursive signatures.
• Dual Ground-Truth Verdicts: Each case provides both strict_verdict (zero-variance matching) and tolerance_verdict (standard AP tolerance policy: ≤ $15.00 or ≤ 1% Prepaid & Add freight, valid partial-delivery billing allowed, missing receipts held, duplicate invoices blocked).

---

Quickstart with Hugging Face datasets

import json
from datasets import load_dataset
from huggingface_hub import snapshot_download

# 1. Load structured case-level benchmark splits (default config: 'cases')
cases = load_dataset("DevLCodes/three-way-matching-dataset", "cases")
print(cases)

sample = cases["test"][0]
print(sample["case_id"], sample["scenario"], "->", sample["tolerance_verdict"])
print("Files:", sample["file_paths"])
print("Discrepancies:", json.loads(sample["discrepancies_json"]))

# 2. Load document-level index ('documents' config — all 7,500 files)
docs = load_dataset("DevLCodes/three-way-matching-dataset", "documents")

# 3. Download raw DOCX, PDF, PNG, and JPG files locally for your agent
repo_dir = snapshot_download("DevLCodes/three-way-matching-dataset", repo_type="dataset")

---

Scenario Taxonomy (2,500 Cases)

Scenario	Cases	Share
damaged_rejected_goods	150	6.0%
duplicate_invoice	125	5.0%
exact_match	550	22.0%
invoice_math_or_tax_error	150	6.0%
missing_warehouse_receipt	125	5.0%
over_shipment_billed	150	6.0%
short_shipment_overbilled	250	10.0%
short_shipment_properly_billed	150	6.0%
unapproved_freight_or_surcharge	175	7.0%
unit_price_hike	250	10.0%
unordered_line_item	125	5.0%
within_tolerance_match	200	8.0%
wrong_po_or_vendor_mismatch	100	4.0%

Scenario Descriptions

1. exact_match: PO, Warehouse Receipt, and Vendor Invoice align 100% on SKUs, quantities, unit prices, tax, and totals (includes realistic vendor description phrasing & date-format variations).
2. within_tolerance_match: Small FOB Origin, Prepaid & Add shipping charge (≤ $12.00) within the $15.00 / 1% AP tolerance threshold (strict_verdict = HOLD_VARIANCE, tolerance_verdict = APPROVED_WITHIN_TOLERANCE).
3. short_shipment_overbilled: Warehouse received fewer units than ordered (with dock clerk note), yet Vendor Invoice billed the full PO quantity.
4. short_shipment_properly_billed: Partial shipment received at dock and Vendor Invoice accurately bills only the delivered quantity (APPROVED_PARTIAL_DELIVERY).
5. over_shipment_billed: Vendor shipped and billed >15–35% extra units beyond PO authorization without a PO amendment.
6. damaged_rejected_goods: Full quantity arrived at dock, but damaged units were rejected during inspection (qty_accepted < qty_received) while Invoice bills full quantity.
7. unit_price_hike: Invoice unit price exceeds contracted PO unit price by +8% to +22%.
8. unapproved_freight_or_surcharge: PO terms specify FOB Destination (Freight Included), but Invoice adds unauthorized freight/surcharge ($185–$550).
9. invoice_math_or_tax_error: Explicit line-extension (qty * unit_price != line_total) or header summation arithmetic error on the Invoice.
10. unordered_line_item: Invoice bills an extra SKU/line item never ordered on the PO or logged on the Warehouse Receipt.
11. missing_warehouse_receipt: Only PO and Invoice are present (2 files); goods receipt is missing/unconfirmed (HOLD_MISSING_RECEIPT).
12. wrong_po_or_vendor_mismatch: Invoice references a non-matching PO number or is issued by a different legal vendor entity / Tax ID than the PO.
13. duplicate_invoice: Packet contains 4 files — PO, Warehouse Receipt, primary Invoice, and a second duplicate/re-billed Invoice against the same shipment (REJECT_DUPLICATE_INVOICE).

---

File Format & Render Mode Distribution (7,500 Documents)

Format	Count	Share
DOCX	2,451	32.7%
JPG	1,346	17.9%
PDF	2,453	32.7%
PNG	1,250	16.7%

Render Mode	Count	Share
openxml_docx	2,451	32.7%
scanned_image_pdf	558	7.4%
scanned_jpg	1,346	17.9%
scanned_png	1,250	16.7%
vector_text_pdf	1,895	25.3%

---

Repository Structure

.
├── README.md                     # Hugging Face Dataset Card & metadata
├── .gitattributes                # Git LFS configuration for binary docs/images
├── ground_truth.json             # Complete nested dataset & AP tolerance policy JSON
├── manifest.csv                  # Case-level summary spreadsheet (2,500 rows)
├── documents_index.csv           # Document-level index (7,500 rows + SHA-256)
├── data/
│   ├── cases_train.jsonl         # 2,000 training cases
│   ├── cases_test.jsonl          # 500 evaluation cases
│   ├── documents_train.jsonl     # 6,000 training document records
│   └── documents_test.jsonl      # 1,500 evaluation document records
└── documents/
    ├── CASE_0001/                # Mixed DOCX + PDF + PNG/JPG packet per case
    ...
    └── CASE_2500/

License

Released under the MIT License. All companies, addresses, tax IDs, SKUs, and transactions are 100% synthetic.