makeitup

Generate synthetic datasets for ML training using LLM. Describe your columns in plain English and get realistic data back.

from makeitup import make

df = make(
    columns={
        "name": "Person's full name",
        "age": "Age between 25 and 55",
        "email": "Work email address",
    },
    num_rows=100
)

Features

• Plain English columns - Describe what you want, get realistic data back
• ML-ready datasets - Add target columns for classification or regression
• Data quality testing - Inject nulls, outliers, typos, or duplicates to test your pipelines
• Multiple formats - Export to CSV, JSON, Parquet, or Excel
• Local model support - Works with OpenAI and any OpenAI-compatible API that supports structured output

Installation

pip install makeitup

Set your OpenAI API key:

export OPENAI_API_KEY=your-api-key

Or create a .env file in your project with OPENAI_API_KEY=your-api-key.

Using a Local Model

makeitup uses structured output to ensure reliable data generation. Local models must support OpenAI-compatible structured output (JSON schema enforcement).

Supported local setups:

• llama.cpp with function calling enabled (llama-server, LM Studio)
• vLLM with --enable-auto-tool-choice
• Ollama (version 0.3.0+) - newer models like llama3.1, qwen2.5
• Any OpenAI-compatible API that implements structured output

Example configuration:

export LLM_BASE_URL=http://localhost:11434/v1  # Ollama
export LLM_MODEL=llama3.1
export LLM_API_KEY=not-needed  # Required by some local servers

Note: Not all local models support structured output. If you encounter errors, verify your model and server support JSON schema enforcement.

Examples

Basic Data

from makeitup import make

# Customer data
df = make(
    columns={
        "customer_id": "Unique customer identifier",
        "name": "Customer full name",
        "email": "Email address",
        "signup_date": "Date when customer signed up, 2020-2024",
    },
    num_rows=100
)

ML Dataset with Target Column

df = make(
    columns={
        "tenure_months": "Months as customer, 1-60",
        "monthly_spend": "Monthly spending in USD, 10-500",
        "support_tickets": "Number of support tickets, 0-10",
    },
    target={
        "name": "churned",
        "prompt": "Boolean indicating if customer churned"
    },
    num_rows=500
)

Data Quality Degradation

# Generate dataset with intentional quality issues for testing data pipelines
df = make(
    columns={
        "name": "Person's full name",
        "age": "Age between 20 and 60",
        "salary": "Annual salary in USD, 30000-150000",
    },
    num_rows=100,
    quality_issues=["nulls", "outliers"],  # Options: nulls, outliers, typos, duplicates
)

Save to File

# CSV, JSON, Parquet, or Excel - format detected from extension
df = make(
    columns={"name": "Product name", "price": "Price in USD, 10-1000"},
    num_rows=200,
    output_path="products.csv"
)

Output Formats

Format	Extension
CSV	.csv
JSON	.json
Parquet	.parquet
Excel	.xlsx

Requirements

• Python >= 3.12
• OpenAI API key or a local model that supports structured output (see "Using a Local Model" above)

Documentation

See DEVELOPER.md for technical details, API reference, and development setup.

License

See LICENSE file for details.