Create test data with AI or a sample file

Generate a dataset with AI

1. In the Data Library, click New dataset, then Generate with AI.
2. Describe the dataset you want in plain language.
3. Set the row count. User defined is selected by default with 30 rows. Enter an exact number from 1 to 500, or choose Let AI decide.
4. Click Generate and wait for the preview to load.
5. Check the suggested dataset name (you can edit it), the columns and the sample records. The preview is read-only. Select the X button in the preview bar to discard it.
6. Click Create new dataset. To change the result instead, click the collapsed prompt to edit it (this discards the preview), then click Generate again.

The new dataset appears in the Data Library.

From your prompt, the AI works out column names and data types, realistic value ranges (sensible dates and prices), relationships between columns (defects rise with temperature, sales peak in holidays) and industry patterns (financial quarters, manufacturing shifts, appointment types).

When to use AI-generated data

Write prompts that generate useful data

Tip: Start with a broad prompt to get the basic structure, then regenerate with more specific constraints.

Example prompts and results

Manufacturing. Prompt:

Generate production line data with timestamp, line ID, units produced, 
defects found, operator name, and temperature readings for 8-hour shifts

Result: 30 rows; average, minimum and maximum temperature per shift (°C); 8 hours per record; realistic operator names rotating across records; temperatures of 20–26°C with natural variation; a shift-level production summary.

Financial services. Prompt:

Create monthly revenue data by product category and region for 2024, 
including columns for month, category, region, revenue, and growth rate

Result: 120 rows (an AI-chosen size); columns Month, Category, Region, Revenue and Growth Rate; months of 2024 in YYYY-MM format; revenue with decimals; growth rates such as -0.05 to 0.05; categories and regions chosen from the prompt.

Healthcare. Prompt:

Generate patient appointment data including appointment ID, patient demographics, 
appointment type, provider name, insurance type, and appointment date/time

Result: sequential appointment IDs (APT-2024-0001); a realistic age and gender spread; types such as Physical, Specialist, Lab Work and Follow-up; providers with matching specialties; insurance such as Medicare, Blue Cross, Aetna and Self-Pay; times in business hours.

E-commerce. Prompt:

Create customer transaction data with order ID, customer name, product, 
quantity, price, order date, and shipping address for online retail

Result: order IDs in ORD-20240001 format; realistic customer names (no actual people); electronics, clothing and home goods; 1-5 items per order; prices of $15-$300 matched to category; recent dates with weekend and holiday patterns; realistic US addresses (no real locations).

Refine a first generation

Check the first result: do the column names, types and value ranges make sense, and are the columns related the way they should be? Then refine:

Original: "Generate customer data"
Refined: "Generate customer data with ages 18-75, email addresses, 
phone numbers, and purchase amounts between $25-$500"

Original: "Include sales rep and revenue"
Refined: "Include sales rep and revenue, where senior reps 
(5+ years experience) have 30% higher average deals"

Original: "Generate order data"
Refined: "Generate order data including some cancelled orders (10%), 
partial refunds, and missing shipping addresses for pickup orders"

Regenerate when column names don't match your expected schema, values look wrong for your industry, or important relationships or business rules are missing. Accept when the structure is close enough for testing and the goal is pipeline testing, not analysis.

Swap generated data for real data later

Generated datasets work as placeholders in a generate, build, replace workflow:

1. Generate sample data that matches the schema you expect.
2. Build the transformation pipeline on it.
3. Test the logic and business rules.
4. Validate the output format and dashboard fit.
5. Replace it with a connector or a file upload when real data arrives.

To keep the schema close to production, name exact columns, formats and relationships in the prompt:

Good prompt: "Include columns named customer_id, order_date, 
product_sku, quantity, unit_price"

Good prompt: "Use YYYY-MM-DD date format, decimal prices 
with 2 decimal places, integer quantities"

Good prompt: "Each customer should have 1-5 orders, 
each order should have 1-3 line items"

See Add or replace data in a dataset for replacing the rows.

What AI-generated data can't do

• Row count: with User defined, the maximum is 500 rows per generation (see Limits and quotas). That suits testing, not analysis.
• Generic patterns: distributions are realistic but not specific to your business.
• Static: no refresh. Generate again if you need new data.
• Approximate schema: it follows general patterns but won't exactly copy your production schema.

Troubleshooting AI-generated data

Load a built-in sample dataset

1. Click New dataset in your project.
2. Select Sample Files.
3. Choose the All samples tab, or one of the Business, Technology, Finance & Real Estate, or Healthcare & Entertainment tabs, and read the descriptions.
4. Click the card of the dataset you want. The card is highlighted; there is no data preview.
5. Click Create new dataset. The dataset appears in the project, ready to explore.

Sample files need no preparation, hold no sensitive data, and give a team identical data to practise on. They are also useful for demos with familiar patterns.

Sample datasets you can load

Each sample is a CSV file of 25 to 63 rows, so it loads quickly and is easy to inspect.

Business samples

Technology samples

Finance & Real Estate samples

Healthcare & Entertainment samples

Pick the right sample for your goal

Tip: Choose sample data that matches your team's domain, for example a manufacturing quality dataset for an operations team.

Practise with a sample dataset

1. Check data quality with Explore Cards: distributions, missing values, outliers and inconsistencies.
2. Try several transformation types: cleaning (standardize company names), filtering (remove test data, focus on a region or period), aggregating (totals, averages, growth rates) and joining.
3. Build an example pipeline: monthly sales reports (Multi-Store Retail), customer lifetime value (SaaS Business Metrics), campaign ROI (Social Media Campaign Performance) or quality tracking (Manufacturing Quality Control).

A first session takes about 15 minutes: pick Multi-Store Retail Sales (2 minutes), explore it with Explore Cards (5 minutes), build a filter pipeline (5 minutes), and generate a sales-by-region dashboard (3 minutes).

Tip: Build your demo pipeline on sample data before a client meeting. You can then repeat the same transformations on their actual data.

What sample files are not

• They aren't your business data or production-ready datasets. They are learning tools and starting points.
• They are smaller than production data, may not include every edge case, and are clean enough to focus on learning Mammoth rather than cleaning.

Move to your own data when you are comfortable with basic transformations, need to solve a specific business problem, or want to build production dashboards. Note which transformations worked on the sample, and start with a subset of your real data.

Related

• Upload files: bring in your own CSV or Excel files.
• Connect a database or app: connect your real systems.
• Add or replace data in a dataset: swap generated rows for real ones.
• Explore cards: explore the new dataset.
• How the pipeline works: build your first pipeline.
• Create a dashboard: visualize the results.