Datasets:

RegalFire
/

dataset-factory

case_id stringclasses 10 values	synthetic bool 1 class	scenario_type stringclasses 7 values	difficulty stringclasses 3 values	amount_tolerance stringclasses 1 value	source_a_amount stringclasses 10 values	source_b_amount stringclasses 9 values	source_a_currency stringclasses 3 values	source_b_currency stringclasses 4 values	source_a_reference stringclasses 10 values	source_b_reference stringclasses 9 values	expected_status stringclasses 6 values	reason_codes listlengths 1 1	human_review_required bool 2 classes
RECON-0000001	true	currency_mismatch	medium	0.01	2460.42	2460.42	EUR	GBP	REF-000001-908	REF-000001-908	REVIEW	[ "CURRENCY_MISMATCH" ]	true
RECON-0000002	true	amount_mismatch	medium	0.01	212.78	221.78	EUR	EUR	INV-000002-631	INV-000002-631	REVIEW	[ "AMOUNT_MISMATCH" ]	true
RECON-0000003	true	rounding_tolerance	medium	0.01	4590.56	4590.57	EUR	EUR	TX-000003-401	TX-000003-401	MATCH_WITH_TOLERANCE	[ "AMOUNT_WITHIN_TOLERANCE" ]	false
RECON-0000004	true	malformed_amount	hard	0.01	4958.96	--100	USD	USD	TX-000004-950	TX-000004-950	INVALID	[ "MALFORMED_AMOUNT" ]	true
RECON-0000005	true	exact_match	easy	0.01	3932.35	3932.35	EUR	EUR	INV-000005-108	INV-000005-108	MATCH	[ "EXACT_MATCH" ]	false
RECON-0000006	true	missing_in_source_b	easy	0.01	1716.35	null	USD	null	PAY-000006-494	null	UNMATCHED_SOURCE_A	[ "MISSING_SOURCE_B" ]	false
RECON-0000007	true	exact_match	easy	0.01	4507.47	4507.47	USD	USD	INV-000007-217	INV-000007-217	MATCH	[ "EXACT_MATCH" ]	false
RECON-0000008	true	rounding_tolerance	medium	0.01	4696.41	4696.40	EUR	EUR	INV-000008-246	INV-000008-246	MATCH_WITH_TOLERANCE	[ "AMOUNT_WITHIN_TOLERANCE" ]	false
RECON-0000009	true	rounding_tolerance	medium	0.01	4722.35	4722.36	EUR	EUR	INV-000009-693	INV-000009-693	MATCH_WITH_TOLERANCE	[ "AMOUNT_WITHIN_TOLERANCE" ]	false
RECON-0000010	true	reference_normalization	medium	0.01	2902.50	2902.50	CHF	CHF	INV-000010-154	inv 000010 154	MATCH_WITH_NORMALIZATION	[ "REFERENCE_FORMAT_VARIATION" ]	false

Dataset Factory / ReconBench

Synthetic data pipelines with explicit acceptance criteria and inspectable QA evidence.

Overview

This portfolio publishes 10 abridged synthetic records, not the commercial 1,000-case dataset. ReconBench targets transaction reconciliation, exception handling and matching-engine regression tests. The public schema is a flattened projection for easy inspection; it cannot reproduce the full benchmark.

Key capabilities

Explicit expected status, reason codes and human-review flags.
A documented 1,000-case source release covering 16 scenarios and 9 statuses.
Reconciliation policies including amount tolerance and reference normalization.
Separate QA evidence and small public samples.
A local tabular engine also documents schema-driven generation and format validation; that generator is not included.

Architecture

flowchart LR
 S[Requirements / scenario definitions] --> G[Private generation]
 G --> V[Deterministic validation]
 V --> P[Accepted records]
 V --> R[Rejected records]
 P --> Q[QA and distributions]
 P --> M[10-row sanitized public projection]

Technology stack

Python, JSONL, deterministic validation, QA reports. Other local Factory V2 outputs include CSV, JSON, Parquet and SQLite-compatible SQL. The local code-repair research pipeline uses model-generated candidates and deterministic checks; its V3 quality gate failed and is excluded from this dataset release.

Example workflow

Define the expected matching policy and scenarios.
Generate synthetic paired records privately.
Validate and separate acceptance from rejection.
Summarize scenario, status and difficulty coverage.
Release only a small inspected projection for portfolio evaluation.

QA evidence / current status

The supplied ReconBench QA v2 report records 1,000 / 1,000 passed, 0 failed, 1,000 unique IDs and 1,000 unique fingerprints. Scenario, status and difficulty totals were cross-checked against release statistics for this portfolio. These are source-dataset checks, not model accuracy, production performance or third-party certification. The original validator was not rerun over the full release during portfolio preparation.

Difficulty: easy 270, medium 410, hard 320. The first 10 source-sample records are a convenience sample, not a balanced or representative evaluation set.

Example inputs / outputs

Scenario	Source A	Source B	Expected
Currency mismatch	2460.42 EUR	2460.42 GBP	REVIEW
Rounding tolerance	4590.56 EUR	4590.57 EUR	MATCH_WITH_TOLERANCE
Malformed amount	4958.96 USD	--100 USD	INVALID

The tolerance example uses a configured threshold of 0.01. Labels come from the original sample.

Screenshots / visual evidence

Three concise portfolio PDFs present the service, QA coverage and selected cases. The dataset viewer displays the actual public rows. No private terminal or desktop capture is included.

Security / privacy design

The records are explicitly synthetic. Account IDs, merchant names, descriptions, timestamps, fingerprints and generation seeds are removed from the public projection. There are no financial customer records or credentials. See SECURITY.md.

What this repository contains

10 abridged JSONL records in the preview split.
Aggregate QA report for the original release, clearly distinguished from the preview.
A small preview-inspection script, architecture documentation and portfolio PDFs.

What remains private/proprietary

Full datasets, generators, prompts, validators, execution logs, reference corpora and commercial delivery packages. No Code Repair V3 candidate is published here.

Author / portfolio

RegalFire · Raven · Clinic Platform · CORTEX

License and intended use

Portfolio evaluation only; see LICENSE. This preview is unsuitable for training or quantitative evaluation. No representativeness, financial correctness outside the supplied scenarios, or production suitability is claimed.

Related applied-AI project

Real Edit Studio — text-to-image and image editing integration using Black Forest Labs FLUX.2 Klein 4B.

Downloads last month: 54

Number of rows:

Total file size:

241 kB