IBM Telco Customer Churn — Synthetic Sample

## Overview A tabular customer churn prediction dataset from a fictional telecommunications company, originally published as part of IBM Sample Datasets and redistributed by scikit-learn on Hugging Face. Each row represents a single customer, with attributes covering demographics, subscribed services, account tenure, billing, and a binary churn label indicating whether the customer left within the last month. Size is in the 1K–10K row range (commonly ~7,043 rows, 21 columns) in CSV format. The dataset is static (snapshot from IBM's sample data, mirrored 2022). ## Schema - customerID — string — unique customer identifier - gender — string — Male / Female - SeniorCitizen — int — 1 if senior citizen, else 0 - Partner — string — Yes / No - Dependents — string — Yes / No - tenure — int — months the customer has stayed with the company - PhoneService — string — Yes / No - InternetService — string — DSL / Fiber optic / No - Contract — string — Month-to-month / One year / Two year - MonthlyCharges — float — current monthly charge in USD - TotalCharges — float — total amount charged over tenure - Churn — string — Yes / No (target variable) - +9 more columns covering OnlineSecurity, OnlineBackup, DeviceProtection, TechSupport, StreamingTV, StreamingMovies, MultipleLines, PaperlessBilling, PaymentMethod ## Sources - Hugging Face: https://huggingface.co/datasets/scikit-learn/churn-prediction — license: CC-BY-4.0 - Originally derived from IBM Sample Data Sets (Telco Customer Churn) ## Methodology The data was synthesized by IBM as a sample/teaching dataset modeling a fictional telecom provider's customer base. Each row encodes a customer's demographic profile, the bundle of services they subscribe to, contract and billing characteristics, and whether they churned in the most recent month. The scikit-learn mirror provides the data as-is in CSV form with no additional transformation beyond hosting on the Hub. ## Known gaps & limitations This is a synthetic/fictional dataset and does not reflect real customer behavior at any actual telecom; patterns learned here may not transfer to production churn modeling. TotalCharges contains a small number of blank strings for zero-tenure customers that need cleaning before numeric use. The dataset is heavily reused in tutorials and Kaggle competitions, so it is effectively memorized by many pretrained models — leakage risk is high if used to evaluate LLMs on tabular reasoning. No timestamp metadata, no geographic detail. Source does not document additional gaps; buyers should validate empirically. ## Intended use & out-of-scope - IS for: classification tutorials, baseline ML benchmarks, feature engineering practice, agent-driven EDA / AutoML demos, RAG over tabular data examples, teaching churn modeling workflows. - NOT for: real-world telco decision-making, evaluating LLMs on unseen tabular tasks (high contamination risk), or fairness audits of real protected populations (data is fictional). Original supplier listing: Telco Customer Churn Prediction (IBM Sample) Classic IBM telco customer churn dataset (~7K rows) with demographics, service subscriptions, account info, and churn label. Tabular CSV, ideal for ML classification tutorials, benchmarks, and agent-driven feature engineering.