VynFi ocel-manufacturing

Regenerated with DataSynth 3.1.1 (2026-04-19). What's new vs. prior releases:

• Behavioral fraud biases now fire on every is_fraud path — weekend ×32, round-dollar ×170, post-close ×3,106 lift measured on fraud-labeled entries (vs ~1× pre-3.1.1).
• Document→JE fraud propagation correctly sets is_fraud_propagated + fraud_source_document_id (was broken in 3.1.0, now verified on every doc-flow JE).
• AML typology coverage reaches the 0.80 evaluator threshold (0.000 → 0.857).
• OCEL timestamps are now microsecond-precision — pandas to_datetime(..., utc=True) retains 100 % of events (was losing 95 %).
• Audit artifacts (audit/audit_opinions.json, audit/key_audit_matters.json) + process_variant_summary.json always ship in the archive.

Configs

Config	Records
events	37,285
objects	9,713
journal_entries	90,834
anomaly_labels	112

Fraud breakdown (DS 3.1.1)

• Total fraud-labeled JEs: 222
• Scheme-propagated (from fraudulent source documents): 18 (8.1%)
• Direct injection (line-level anomaly): 204

is_fraud_propagated is set by DataSynth 3.1+ when a fraudulent source document fans out to its derived journal entries. Use this to split ring-level (cross-document scheme) from slip-level (isolated anomaly) fraud populations when training detection models.

Quick start

from datasets import load_dataset

ds = load_dataset("VynFi/ocel-manufacturing", name="events", split="train")
print(ds.features)
print(ds[0])

Or via the VynFi Python SDK (v1.5.1):

import os
from vynfi import VynFi

client = VynFi(api_key=os.environ["VYNFI_API_KEY"])
job = client.jobs.generate_config(config={...})  # see https://github.com/VynFi/VynFi-python/tree/main/examples

See the SDK cookbook for worked examples:

• examples/document_level_fraud.py
• examples/behavioral_fraud_patterns.py
• examples/sector_dag_presets.py
• examples/audit_opinions_kam.py

License

Apache 2.0. Entirely synthetic — no real individuals, companies, or transactions.