Seven datasets built for practicing on, not scraped from anywhere. Each one has real structure in it — price genuinely depends on floor area, churn genuinely depends on contract type — so a model you fit actually finds something. One of them is deliberately filthy, because cleaning is the skill nobody teaches.
Load any of them in the notebook with one line. Each CSV loads from this site straight into the Python running in your tab, and nothing you do with it is sent back.
df = pm.load("housing")
df.head()Six Pune neighbourhoods, 600 listings. Price is driven by floor area, bathrooms, age and parking, with a per-area premium on top.
Practice: Linear regression, feature engineering, one-hot encoding, residual plots.
idareasqftbedroomsbathroomsage_yearsparkingfloorprice_inrdf = pm.load("housing")Two years of daily readings for four cities with genuinely different climates, including a monsoon season that shows up in the rainfall.
Practice: Datetime parsing, resampling, groupby, rolling averages, seasonality.
datecitytemp_chumidity_pctrain_mmwind_kphdf = pm.load("weather")1,400 orders across four categories and three channels. Electronics skew to the app, books to the web — a pivot table will find it.
Practice: groupby, pivot_table, revenue aggregation, categorical analysis.
order_iddatecategoryproductregionchannelunitsunit_pricediscount_pctrevenuedf = pm.load("retail")500 invented films with budgets, revenue and ratings. Revenue follows budget loosely, with a long tail of flops and outsized hits.
Practice: Filtering, correlation, log transforms, outlier hunting, plotting.
titleyeargenreruntime_minbudget_musdrevenue_musdratingvotesdf = pm.load("movies")400 rows carrying six kinds of real-world dirt on purpose: four date formats, inconsistent casing and whitespace, missing values, currency stored as text, impossible ages, and exact duplicate rows.
Practice: Everything cleaning: dropna, fillna, str.strip, str.lower, to_datetime, to_numeric, duplicated, range validation.
user_idemailnamesignup_dateplancountryagespendnewsletterdf = pm.load("signups")800 students. Final score is a real function of study hours, prior score, attendance, sleep and tutoring — plus noise you cannot model.
Practice: Regression, train/test split, classification on the pass/fail column.
student_idstudy_hours_weekattendance_pctprior_scoretutoringsleep_hoursfinal_scoreresultdf = pm.load("students")900 customers. Month-to-month contracts and frequent support calls genuinely predict churn; long tenure genuinely protects against it.
Practice: Binary classification, class imbalance, confusion matrix, feature importance.
customer_idtenure_monthscontractmonthly_chargesupport_callspaperless_billingpayment_methodchurneddf = pm.load("churn")scikit-learn ships several classic datasets inside the package itself, so they work here with no network at all. Those are not duplicated above — reach for them directly:
from sklearn.datasets import load_iris, load_wine, load_diabetes, load_breast_cancer
X, y = load_iris(return_X_y=True, as_frame=True)Every file above is generated from a seeded random number generator. The rows are invented: no real person, listing, order or customer is represented, and nothing here carries a licence you need to worry about.