---
license: mit
task_categories:
- tabular-classification
- text-classification
language:
- ar
- en
tags:
- synthetic-data
- gcc
- invoices
- ecommerce
- erp
- fintech
- testing
- data-quality
- arabic
- benchmark
pretty_name: GCC-InvoiceMath-Verified
size_categories:
- n<1K
---

# GCC-InvoiceMath-Verified

A deterministic synthetic benchmark for testing invoice arithmetic, numeric parsing, rounding, and claimed-total validation across six GCC currencies.

## Overview

GCC-InvoiceMath-Verified contains **600 records / 300 causal pairs**.

Each pair contains:

- one mathematically correct invoice
- one counterfactual invoice where exactly one claimed total differs by one minor currency unit

The dataset is intended for:

- ERP QA
- e-invoicing integration testing
- financial software regression testing
- structured-output validation
- numerical reasoning evaluation
- synthetic data benchmarking

This benchmark is **not a tax-law compliance benchmark**.

It does not certify compliance with ZATCA, UAE e-invoicing rules, PINT, UBL, or any national tax regulation.

## Countries and currencies

| Country | Currency | Records |
|---|---|---:|
| Saudi Arabia | SAR | 100 |
| United Arab Emirates | AED | 100 |
| Qatar | QAR | 100 |
| Kuwait | KWD | 100 |
| Bahrain | BHD | 100 |
| Oman | OMR | 100 |

SAR, AED and QAR use two benchmark fractional digits.

KWD, BHD and OMR use three benchmark fractional digits.

## Scenario families

The benchmark contains 10 deterministic scenario families:

1. `half_up_tie`
2. `line_vs_document_rounding`
3. `line_allowance`
4. `line_charge`
5. `prepayment`
6. `credit_note`
7. `large_amount_stress`
8. `zero_rate`
9. `mixed_rates`
10. `zero_due`

Each family contains **60 records**.

## Numeric rendering

Four rendering styles are equally represented:

- ASCII
- grouped ASCII
- Arabic-Indic digits
- grouped Arabic-Indic digits

Each style contains **150 records**.

This makes it possible to test parsers handling both Western and Arabic-Indic numeric representations.

## Dataset structure

Important fields include:

```json
{
  "id": "...",
  "pair_id": "...",
  "split": "train",
  "country": "SA",
  "family": "half_up_tie",
  "contract_id": "gcc-invoice-math-1.0",
  "invoice": {
    "currency": "SAR",
    "document_type": "invoice",
    "numeric_style": "arabic_indic",
    "lines": [],
    "prepaid": "...",
    "claimed_totals": {}
  },
  "expected": {
    "valid": true,
    "totals": {},
    "error_fields": []
  },
  "mutation": null,
  "fingerprint_sha256": "..."
}
Ground truth
Ground truth is generated procedurally using Python Decimal arithmetic with explicit ROUND_HALF_UP behavior.
Validation is performed independently using fractions.Fraction.
The validator does not import generator code.
This provides two different arithmetic implementations:
- generator oracle: Decimal
- validation oracle: rational arithmetic
Validation results
- Records generated: 600
- Records validated: 600
- Pass rate: 100%
- Valid examples: 300
- Invalid examples: 300
- Unique causal pairs: 300
The release also checks:
- JSON Schema compliance
- unique IDs
- unique semantic source invoices
- fingerprints
- country/currency consistency
- pair integrity
- train/validation/test pair isolation
- exactly one causal mutation per invalid record
- exactly one minor-unit difference
- generator/validator oracle agreement
Splits
- train: 360
- validation: 120
- test: 120
Positive and negative members of a causal pair are always kept in the same split.
Synthetic provenance
All records are procedurally generated.
- no real invoices
- no PII
- no copied commercial records
- no real tax identifiers
- no customer transaction data
External sources were used only for background research and currency-precision verification.
Important limitations
This dataset does not cover:
- XML / UBL validation
- digital signatures
- QR codes
- live tax submission
- regulatory certification
- real tax IDs
- OCR images
- exchange rates
- payment reconciliation
- jurisdiction-specific VAT interpretation
Tax rates used in the benchmark are synthetic testing parameters.
Intended use
Recommended uses:
- unit tests
- regression tests
- ERP adapter QA
- invoice parser evaluation
- financial-agent evaluation
- data-quality pipelines
- numeric-reasoning benchmarks
Commercial customization
Need a private benchmark matching your ERP, invoice schema, accounting rules, API, or failure modes?
Custom synthetic datasets can be generated with:
- client-specific field mappings
- custom currencies
- custom rounding behavior
- additional failure modes
- private held-out regression suites
- custom JSON / CSV / Parquet / SQL formats
- validation scripts
- QA reports
The commercial value is in adapting the benchmark to a real system, not simply generating more generic rows.
Reproducibility
The generator is deterministic and seed-based.
Default seed:
20260930
See generator.py and validator.py.
License
MIT License for the original benchmark code and synthetic fixtures.
See LICENSE.

---

## RegalFire — Custom / Private Dataset Work

RegalFire builds custom AI datasets, evaluation sets and data pipelines for:

- AI agents
- computer-use systems
- multimodal models
- world models
- RAG systems
- code agents
- enterprise AI

Available services include:

- synthetic data generation
- private evaluation datasets
- agent trajectories
- failure / recovery datasets
- multimodal RGB / segmentation / state-action data
- web data acquisition
- cleaning and deduplication
- structured dataset packaging
- continuous dataset production

For custom or private work:

**Email: ootiris@gmail.com**

Hugging Face: **RegalFire**