arun-gharami/lead-ai-fraud-detection-dataset

📊 Lead.AI Fraud Detection Dataset

5,000-Row Synthetic Tabular Benchmark — Ready to Train, Ready to Publish

 

Published by Lead.AI Labs · Author: Arun Kumar Gharami

What This Dataset Is For

A clean, Parquet-formatted, immediately loadable synthetic fraud detection dataset built for researchers, ML engineers, and course instructors who need realistic tabular financial data without the legal complexity of real transaction data.

Use it to:

• —Build and benchmark fraud classifiers in hours, not weeks
• —Run XAI / SHAP experiments on financial features
• —Teach imbalanced classification in courses or workshops
• —Prototype a fraud detection proof-of-concept for a client demo

⚠️ Synthetic data. No real customers, no real transactions, no PII. Safe to use, share, and publish.

Dataset at a Glance

Data Fields

Load in 3 Lines

from datasets import load_dataset

ds = load_dataset("arun-gharami/lead-ai-fraud-detection-dataset")
df = ds["train"].to_pandas()

With pandas directly

import pandas as pd

df = pd.read_parquet(
    "hf://datasets/arun-gharami/lead-ai-fraud-detection-dataset/data/train-00000-of-00001.parquet"
)
print(df["risk_label"].value_counts())

With DuckDB (fast SQL on Parquet)

import duckdb

duckdb.query("""
    SELECT risk_label, COUNT(*) as n, ROUND(AVG(transaction_amount), 2) as avg_amount
    FROM read_parquet('hf://datasets/arun-gharami/lead-ai-fraud-detection-dataset/data/train-00000-of-00001.parquet')
    GROUP BY risk_label
""").df()

Ready-to-Use Model Trained on This Data

Don't want to train your own? The Lead.AI Fraud Shield is already trained and ready to load:

import joblib, pandas as pd
model = joblib.load("model/model.joblib")   # from lead-ai-fraud-shield repo

Need More Data?

This dataset has 5K rows and 14 features — good for fast prototyping and course projects.

For production model training or research publication, use [Dataset v2 → 100K rows, 21 features](https://huggingface.co/datasets/arun-gharami/lead-ai-fraud-detection-dataset-v2) which adds transaction type, day-of-week, geographic region, customer risk score, and more.

Bias & Fairness Note

Synthetically generated. Country and device features may encode assumptions that don't reflect real fraud distributions. Audit for proxy discrimination before using a model trained here in any context involving real individuals.

Privacy

No PII. All IDs, amounts, and behavioral features are synthetically generated.

Citation

@misc{gharami2024frauddataset,
  author       = {Arun Kumar Gharami},
  title        = {Lead.AI Fraud Detection Dataset: Synthetic Tabular Benchmark for XAI Research},
  year         = {2024},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/arun-gharami/lead-ai-fraud-detection-dataset}}
}

Lead.AI Labs — Trustworthy AI Systems for Practical Business Intelligence lead-ai.us · LinkedIn · GitHub