dataxid 0.3.0

The Synthetic Data API — privacy-preserving synthetic data generation

DataXID Python SDK

High-fidelity synthetic data generation for single-table, multi-table, and time series data.

Installation

pip install dataxid

Quick Start

import dataxid
import pandas as pd

dataxid.api_key = "dx_..."
dataxid.enable_logging("info")  # optional: see training progress

df = pd.read_csv("data.csv")
synthetic = dataxid.synthesize(data=df, n_samples=1000)

Multi-Table & Time Series

Synthesize related tables with referential integrity. Child tables are generated sequentially by default — preserving realistic per-entity patterns like transaction counts, temporal ordering, and sequence lengths.

from dataxid import Table

accounts = Table(pd.read_csv("accounts.csv"), primary_key="account_id")
transactions = Table(pd.read_csv("transactions.csv"),
                     foreign_keys={"account_id": accounts})

synthetic = dataxid.synthesize_tables({
    "accounts": accounts,
    "transactions": transactions,
})

synthetic["accounts"]       # synthetic accounts with auto-assigned PKs
synthetic["transactions"]   # sequential transactions per account, valid FKs

Per-table generation controls are passed as dicts keyed by table name:

from dataxid import Synthetic, Distribution

accounts_preset = Synthetic(n=1000)
transactions_preset = Synthetic(n=5000, seed=42)

country_distribution = Distribution(
    column="country",
    probabilities={"US": 0.6, "UK": 0.4},
)

synthetic = dataxid.synthesize_tables(
    tables={"accounts": accounts, "transactions": transactions},
    synthetic={
        "accounts": accounts_preset,
        "transactions": transactions_preset,
    },
    distribution={"accounts": country_distribution},
)

Iterative workflow

When you want to train once and generate many times — for example, running several sampling strategies against the same model — split the call into Model.create and model.generate:

model = dataxid.Model.create(data=df)
synthetic_a = model.generate(n_samples=1000, diversity=0.8)
synthetic_b = model.generate(n_samples=1000, diversity=1.2)
model.delete()

How It Works

DataXID is built on a privacy-by-architecture principle. Data encoding and decoding happen entirely on your machine; only abstract embeddings are shared with the API for model training. Raw data never leaves your environment.

Configuration

Parameter	Default	Description
embedding_dim	64	Embedding size (larger = more expressive)
model_size	"medium"	Model capacity: "small", "medium", "large"
max_epochs	100	Maximum training epochs
batch_size	256	Training batch size
privacy	Privacy()	Privacy config (see below)

config = dataxid.ModelConfig(
    embedding_dim=128,
    model_size="large",
    max_epochs=50,
    privacy=dataxid.Privacy(enabled=True, noise=0.2),
)
model = dataxid.Model.create(data=df, config=config)

A plain dict is also accepted for quick experiments: config={"embedding_dim": 128}.

Privacy

Privacy settings are grouped under a dedicated Privacy config:

Parameter	Default	Description
enabled	False	Add Gaussian noise to embeddings before they leave the machine
noise	0.1	Noise scale (Gaussian std) when enabled=True
protect_rare	True	Hide rare categorical values behind a <protected> token
rare_strategy	"mask"	How protected values appear: "mask" or "sample"

Advanced Generation

Rebalance category distributions

Override the natural distribution of a categorical column:

model = dataxid.Model.create(data=df)

distribution = dataxid.Distribution(
    column="gender",
    probabilities={"M": 0.5, "F": 0.5},
)
synthetic = model.generate(n_samples=1000, distribution=distribution)

Bias correction

Reduce statistical parity gaps across sensitive attributes:

bias = dataxid.Bias(
    target="income",
    sensitive=["gender", "race"],
)
synthetic = model.generate(n_samples=1000, bias=bias)

Conditional generation

Fix known values and let the model complete the rest:

conditions = pd.DataFrame({"income": [">50K"] * 1000})

synthetic = model.generate(conditions=conditions)

Impute missing values

Fill NaN cells with model predictions; non-NULL cells are preserved:

model = dataxid.Model.create(data=df)
filled = model.impute(df, trials=3, pick="mode")

Tuning presets

Bundle generation-time knobs into a reusable preset:

preset = dataxid.Synthetic(
    n=1000,
    seed=42,
    diversity=0.8,
    rare_cutoff=0.95,
)
synthetic = model.generate(synthetic=preset)

Direct keyword arguments override preset values:

synthetic = model.generate(synthetic=preset, diversity=1.2)  # diversity=1.2 wins

Logging

dataxid.enable_logging("info")   # see training progress, epoch stats
dataxid.enable_logging("debug")  # verbose: includes HTTP requests
dataxid.disable_logging()        # turn off (default state)

Or via environment variable (no code change needed):

DATAXID_LOG=info python my_script.py

Error Handling

import dataxid

try:
    synthetic = dataxid.synthesize(data=df)
except dataxid.AuthenticationError:
    print("Invalid API key")
except dataxid.QuotaExceededError as e:
    print(f"Quota exceeded. Upgrade: {e.upgrade_url}")
except dataxid.RateLimitError as e:
    print(f"Rate limited. Retry after: {e.retry_after}s")
except dataxid.DataxidError as e:
    print(f"Error: {e}")

Links

• Documentation
• API Reference
• GitHub
• Examples