Mimic Mode

Mimic mode takes a real CSV file (or DataFrame), analyzes every column's statistical fingerprint, and produces a fresh synthetic dataset that matches the original's structure — without reusing a single real value.

It's the fastest path from "I have sensitive production data" to "I have safe, shareable synthetic data".

Quickstart

import misata

# From a CSV file
tables = misata.mimic("customers.csv")

# Scale to a different size
tables = misata.mimic("customers.csv", rows=100_000)

# From a DataFrame you already have
import pandas as pd
df = pd.read_csv("orders.csv")
synthetic = misata.mimic(df, rows=50_000)

CLI

# Same number of rows as the source
misata mimic customers.csv

# Scale up to 100 k rows, write to ./synthetic/
misata mimic orders.csv --rows 100000 --output ./synthetic

# Reproducible output
misata mimic data.csv --seed 42

Multi-table

Pass a list of files. Each becomes its own synthetic table, named after the file stem:

tables = misata.mimic(["customers.csv", "orders.csv", "products.csv"])
# tables["customers"], tables["orders"], tables["products"]

How it works

For each column, Misata's DataProfiler runs a five-step analysis:

Step	What it does
Type detection	Identifies boolean, date, integer, float, or text
Distribution fitting	Fits lognormal (right-skewed), normal, or uniform — whichever matches the data
Cardinality check	Low-cardinality columns become categoricals with real frequency weights
Semantic inference	Detects email, name, city, country, latitude, URL, phone, etc.
Range capture	Records min/max for numerics, start/end for dates

None of the original values are stored or emitted. The profiler only retains statistical parameters.

Distribution fitting logic

if all values > 0 AND skew > 1.0:
    → lognormal (mu, sigma from log-space moments)
elif cardinality < 5% of rows AND unique values < 200:
    → categorical with observed frequencies
else:
    → normal (mean, std)

Supported column types

Source column	What Misata generates
email	Realistic email addresses (never real ones)
first_name / last_name	Diverse synthetic names
city, country, state	Real place names from vocab
latitude / longitude	Coordinates clustered around real cities
postal_code / zip	Format-correct postal codes
Dates	Dates within the observed range
Booleans	Same true/false ratio
Low-cardinality text	Category distribution preserved
High-cardinality text	Semantic type inferred and regenerated

Reproducibility

# Same seed → identical output every run
tables = misata.mimic("customers.csv", rows=10_000, seed=42)

When to use mimic vs. generate

Situation	Use
You have a real schema to copy	mimic()
You want to describe a dataset from scratch	generate()
You need relational FK integrity across tables	generate() with relationships
You need to match an existing DB's distributions	mimic() with multiple CSVs

API reference

misata.mimic(
    source,          # str | Path | pd.DataFrame | list of those
    rows=None,       # int — defaults to same count as source
    seed=None,       # int — for reproducibility
    table_name="table",  # str — used when source is a DataFrame
) -> dict[str, pd.DataFrame]