遇见数据集

VynFi/audit-p2p

收藏
Hugging Face2026-04-19 更新2026-04-26 收录
官方服务:

资源简介:

--- license: apache-2.0 task_categories: - tabular-classification tags: - synthetic - financial-data - vynfi - audit - procure-to-pay - fraud-detection size_categories: - 10K<n<100K configs: - config_name: purchase_orders data_files: "purchase_orders/*.parquet" - config_name: goods_receipts data_files: "goods_receipts/*.parquet" - config_name: vendor_invoices data_files: "vendor_invoices/*.parquet" - config_name: payments data_files: "payments/*.parquet" - config_name: journal_entries data_files: "journal_entries/*.parquet" --- # VynFi audit-p2p 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 | |--------|---------| | `purchase_orders` | 50 | | `goods_receipts` | 59 | | `vendor_invoices` | 50 | | `payments` | 78 | | `journal_entries` | 113,593 | ## Fraud breakdown (DS 3.1.1) - **Total fraud-labeled JEs:** 260 - **Scheme-propagated** (from fraudulent source documents): 10 (3.8%) - **Direct injection** (line-level anomaly): 250 `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 ```python from datasets import load_dataset ds = load_dataset("VynFi/audit-p2p", name="purchase_orders", split="train") print(ds.features) print(ds[0]) ``` Or via the VynFi Python SDK (v1.5.1): ```python 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.

--- license: Apache-2.0 task_categories: - 表格分类(tabular-classification) tags: - 合成数据集(synthetic) - 金融数据(financial-data) - VynFi - 审计 - 采购到付款(procure-to-pay) - 欺诈检测(fraud-detection) size_categories: - 10000 < 样本量 < 100000 configs: - config_name: 采购订单(purchase_orders) data_files: "purchase_orders/*.parquet" - config_name: 收货单(goods_receipts) data_files: "goods_receipts/*.parquet" - config_name: 供应商发票(vendor_invoices) data_files: "vendor_invoices/*.parquet" - config_name: 付款记录(payments) data_files: "payments/*.parquet" - config_name: 日记账分录(journal_entries) data_files: "journal_entries/*.parquet" --- # VynFi audit-p2p 数据集 本数据集由**DataSynth 3.1.1**(2026-04-19)重新生成。相较于此前版本的更新内容如下: - **行为欺诈偏置现已在所有`is_fraud`路径中生效**——针对标记为欺诈的日记账分录,已测得周末相关欺诈提升32倍、整数金额相关欺诈提升170倍、结账后相关欺诈提升3106倍(而3.1.1版本前该提升约为1倍)。 - **文档→日记账分录(Journal Entries,JE)的欺诈传播逻辑**现已正确设置`is_fraud_propagated`与`fraud_source_document_id`字段(该功能在3.1.0版本中存在缺陷,现已在所有文档流日记账分录中完成验证)。 - **反洗钱(Anti-Money Laundering,AML)类型学覆盖度**已达到0.80的评估阈值(从0.000提升至0.857)。 - **开放流程事件日志(OCEL)时间戳**现已支持微秒级精度——使用pandas的`to_datetime(..., utc=True)`可完整保留100%的事件数据(此前会丢失95%的事件)。 - **审计工件**(`audit/audit_opinions.json`、`audit/key_audit_matters.json`)与`process_variant_summary.json`现已始终随归档包一同发布。 ## 配置项详情 | 配置项名称 | 样本量 | |--------|---------| | 采购订单(purchase_orders) | 50 | | 收货单(goods_receipts) | 59 | | 供应商发票(vendor_invoices) | 50 | | 付款记录(payments) | 78 | | 日记账分录(journal_entries) | 113,593 | ## 欺诈数据分布(DS 3.1.1版本) - **总欺诈标记日记账分录数**:260 - **方案传播型欺诈**(源自欺诈性源文档):10例(占比3.8%) - **直接注入型欺诈**(行级异常):250 当欺诈性源文档扩散至其衍生的日记账分录时,DataSynth 3.1及以上版本会设置`is_fraud_propagated`字段。该字段可用于在训练欺诈检测模型时,将跨文档方案级(环级)欺诈与单文档孤立异常级(单据级)欺诈群体进行拆分。 ## 快速上手 python from datasets import load_dataset ds = load_dataset("VynFi/audit-p2p", name="purchase_orders", split="train") print(ds.features) print(ds[0]) 或通过VynFi Python软件开发工具包(SDK,v1.5.1)使用: python 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 可参考SDK Cookbook获取完整实操示例: - `examples/document_level_fraud.py` - `examples/behavioral_fraud_patterns.py` - `examples/sector_dag_presets.py` - `examples/audit_opinions_kam.py` ## 许可证 Apache 2.0。本数据集完全为合成数据,未使用任何真实个人、企业或交易信息。

提供机构:
VynFi
二维码
社区交流群
二维码
科研交流群
商业服务