VynFi/vynfi-je-network-2k

VynFi JE Network — 2 000-entity group consolidated

Update — v5.35.1 (P0c corpus-scale amounts): regenerated with the P0c amount calibration. The per-line amount median is now corpus-scale, p99/p50 ~200×, Benford MAD ~0.001 — up from the prior ~$300. Structural levers (lines/JE ~3.7, multi-currency, allocation lines) are unchanged. Where older embedded stats below conflict with this note, this note is authoritative.

1 558 254 journal-entry network edges produced by the v5.35.1 group- audit simulation engine over 2 000 entities + consolidated FS bundle. Each row is one debit↔credit edge from a single journal entry, attributed to its source entity, with the document-chain, fraud / anomaly, and intercompany-pair labels carried through from the generator.

Headline

field	value
Total edges	1 558 254
Entities	2 000
Fraud edges (sampled)	≈ 5.6 % (line-level)
IC pairs matched	6 814
IC eliminations	13 628 edges
Generation engine	mivertowski/SyntheticData @ v5.31
Aggregate walltime	65 min end-to-end on a 30-core / 222 GB VM
Aggregate peak RSS	0.382 GB (vs prior-version 218 GB OOM — see Phase 6 + 7 success doc)
Determinism	Reproducible from seed field in manifest

What this dataset is for

• Edge-level fraud detection. Binary or multi-class (fraud_type) edge classification — drop into any GNN framework via from_account → to_account directed edges.
• Account-network / document-chain analytics. Each edge carries predecessor_edge_id so document-flow paths (PO → invoice → payment etc.) reconstruct without joins.
• Intercompany matching / elimination research. 6 814 IC pairs (seller + buyer sides) labelled by ic_pair_id + ic_partner_entity. 13 628 elimination edges with is_eliminated=true and eliminates_ic_pair_id populated.
• Anomaly-injection benchmarking. is_anomaly flags non- fraud anomalies (errors, process deviations, statistical outliers). fraud_type enum carries the typology when is_fraud=true.

Schema (19 columns)

column	type	notes
edge_id	utf8	UUID v5 over (document_id, dr_line, cr_line)
document_id	utf8	Source JE's document UUID
entity_code	utf8	Source entity (one of 2 000)
posting_date	utf8	ISO-8601 date
from_account	utf8	Credit-side GL account
to_account	utf8	Debit-side GL account
from_line_id	utf8	Credit line transaction ID
to_line_id	utf8	Debit line transaction ID
amount	utf8	Decimal (string-encoded to preserve precision)
confidence	float64	1.0 for Method-A bijective; <1.0 for Cartesian
predecessor_edge_id	utf8	Empty if the chain root
business_process	utf8	P2P / O2C / R2R / H2R / A2R / etc.
is_fraud	bool	Fraud-injected JE
is_anomaly	bool	Non-fraud anomaly
fraud_type	utf8 (nullable)	Fine-grained typology when is_fraud
ic_pair_id	utf8 (nullable)	Intercompany pair identifier
ic_partner_entity	utf8 (nullable)	Counterparty entity code
is_eliminated	bool	True on consolidation-elimination edges
eliminates_ic_pair_id	utf8 (nullable)	Pair the elimination cancels

Loading

import pyarrow.parquet as pq
import pandas as pd

# Whole dataset (5.8 GB on disk)
table = pq.read_table("je_network.parquet")
df = table.to_pandas()

# Stream by row group (66 groups, ≈ 1M edges each)
pf = pq.ParquetFile("je_network.parquet")
for i in range(pf.num_row_groups):
    df = pf.read_row_group(i).to_pandas()
    ...

# Edge index for PyTorch Geometric / DGL
edges = df[["from_account", "to_account"]].values
labels = df["is_fraud"].values

How it was produced

1. Group generation (per-entity orchestrator): 2 000 entities running the v5.31 SOTA engine with multi-currency, source- conditional Dirichlet account-pair sampling, top-50 archetype coverage, recurring/standard-journal templates, business-unit dimension, reversal/correction process, blank-source posting (~21 %), and heavy-tail consolidation outliers (B2 / #154, default rate 0.001).
2. Streaming aggregate (#156, v5.31 C1 Phases 6 + 7): each per-entity journal_entries.json streamed through a JeNetworkEdgeBuilder that yields one JE at a time and maintains the line_id → edge_id map across calls for cross-JE predecessor chains.
3. IC matching (#156 Phase 3 consume-by-value matcher) joined seller / buyer sides by ic_pair_id, 100 % coverage.
4. Elimination factory (#156 Phase 7 defensive resolution) posted 13 628 elimination edges using the manifest-driven notional with a 3-tier amount-resolution fallback for any IC JE that an upstream anomaly injector had corrupted.

Methodology notes

• Multi-shard variance: per the v5.31 T3 multi-shard BF study, single-shard Sajja-composite CV is ≈ 25 % (P1 autocorr CV ≈ 103 %, P2 burst-length CV ≈ 1.9 %). Edge-level metrics computed on this single 2k shard are reference points; multi-seed reruns remain methodology-bound.
• Intercompany contract: 4 of the 6 814 IC pairs (0.06 %) hit the elimination defensive-fallback path because an upstream ReversedAmount anomaly had flipped their debit / credit. Phase 7+ (#160) closes the source of this corruption going forward; existing eliminations remain correct via the total-credit fallback.
• Synthetic only: all data is procedurally generated. No client / corpus content; entity names, accounts, and amounts are synthetic.

License

Apache-2.0 — same as the engine.

Citation

@misc{vynfi_je_network_2k,
  title = {VynFi JE Network: 2 000-entity group consolidated edges},
  author = {Ivertowski, Michael},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/VynFi/vynfi-je-network-2k}}
}

Related artefacts

• Engine: mivertowski/SyntheticData
• Sister dataset (1M JE rows): VynFi/vynfi-journal-entries-1m
• GNN model (v5.29 retrain): VynFi/vynfi-je-fraud-gnn