be-great 0.0.14

Generating Realistic Tabular Data using Large Language Models

Generation of Realistic Tabular data

with pretrained Transformer-based language models

Our GReaT framework leverages the power of advanced pretrained Transformer language models to produce high-quality synthetic tabular data. Generate new data samples effortlessly with our user-friendly API in just a few lines of code. Please see our publication for more details.

GReaT framework has also been adopted in practice on Google’s Kaggle platform, where it has been used to generate synthetic datasets across multiple competitions.

 

GReaT Installation

The GReaT framework can be easily installed using with pip - requires a Python version >= 3.9:

pip install be-great

GReaT Quickstart

In the example below, we show how the GReaT approach is used to generate synthetic tabular data for the California Housing dataset.

from be_great import GReaT
from sklearn.datasets import fetch_california_housing

data = fetch_california_housing(as_frame=True).frame

model = GReaT(llm='tabularisai/Qwen3-0.3B-distil', batch_size=32,  epochs=5,
              fp16=True, dataloader_num_workers=4)
model.fit(data)
synthetic_data = model.sample(n_samples=100)

Imputing a sample

GReaT also features an interface to impute, i.e., fill in, missing values in arbitrary combinations. This requires a trained model, for instance one obtained using the code snippet above, and a pd.DataFrame where missing values are set to NaN. A minimal example is provided below:

# test_data: pd.DataFrame with samples from the distribution
# model: GReaT trained on the data distribution that should be imputed

# Drop values randomly from test_data
import numpy as np
for clm in test_data.columns:
    test_data[clm]=test_data[clm].apply(lambda x: (x if np.random.rand() > 0.5 else np.nan))

imputed_data = model.impute(test_data, max_length=200)

Saving and Loading

GReaT provides methods for saving a model checkpoint (besides the checkpoints stored by the huggingface transformers Trainer) and loading the checkpoint again.

model = GReaT(llm='tabularisai/Qwen3-0.3B-distil', batch_size=32,  epochs=5, fp16=True)
model.fit(data)
model.save("my_directory")  # saves a "model.pt" and a "config.json" file
model = GReaT.load_from_dir("my_directory")  # loads the model again

# supports remote file systems via fsspec
model.save("s3://my_bucket")
model = GReaT.load_from_dir("s3://my_bucket")

Optimizing GReaT for Challenging Datasets

When working with small datasets or datasets with many features, GReaT offers specialized parameters to improve generation quality:

# For small datasets or datasets with many features
model = GReaT(
    llm='tabularisai/Qwen3-0.3B-distil',  
    float_precision=3,  # Limit floating-point precision to 3 decimal places
    batch_size=8,       # Use smaller batch size for small datasets
    epochs=100,         # Train for more epochs with small data
    fp16=True           # Enable half-precision training for faster computation and lower memory usage
)
model.fit(data)

# Use guided sampling for higher quality generation with complex feature sets
synthetic_data = model.sample(
    n_samples=100,
    guided_sampling=True,     # Enable feature-by-feature guided generation
    random_feature_order=True,  # Randomize feature order to avoid bias
    temperature=0.7           # Control diversity of generated values, use lower temperature for challenging data
)

The guided_sampling=True parameter enables a feature-by-feature generation approach, which can produce more reliable results for datasets with many features or complex relationships. While potentially slower than the default sampling method, it can help overcome generation challenges with difficult datasets.

The float_precision parameter limits decimal places in numerical values, which can help the model focus on significant patterns rather than memorizing exact values. This is particularly helpful for small datasets where overfitting is a concern.

Conditional Synthetic Data Generation

GReaT supports constrained sampling with logical operators — generate synthetic tabular data that satisfies conditions like age >= 30 or city != 'New York'. Constraints are enforced during token generation, so every output row is valid with zero waste.

from be_great import GReaT
from ucimlrepo import fetch_ucirepo

# Load the UCI Adult (Census Income) dataset
adult = fetch_ucirepo(id=2)
df = adult.data.features[["age", "workclass", "education", "sex", "hours-per-week"]].copy()
df["income"] = adult.data.targets["income"]
df = df[~df.isin(["?"]).any(axis=1)].dropna()

model = GReaT(llm='distilgpt2', epochs=50, batch_size=32, float_precision=0)
model.fit(df)

# Generate synthetic data with constraints
synthetic_data = model.sample(
    n_samples=100,
    conditions={
        "age": ">= 40",
        "hours-per-week": "<= 40",
        "sex": "!= 'Male'",
    },
)

Supported operators for numeric columns: >=, <=, >, <, ==, !=. For categorical columns: ==, != (quote values with single quotes, e.g. "== 'Female'"). Multiple conditions can be combined in a single call. Guided sampling is enabled automatically when conditions are provided.

Mock Data Generation (no real data needed)

Generate synthetic rows from a declarative schema — no fit(), no training data. Useful for privacy-safe dummy data, test fixtures, dev environments, and schema prototyping. Works best when paired with a tabular-pretrained LLM such as tabularisai/Qwen3-0.3B-distil.

from be_great import GReaT
from be_great.great_mock_datasets import save_schema, load_schema

schema = {
    "age":            {"type": "num", "range": (18, 99), "integer": True,
                       "dist": "normal", "mean": 38, "std": 13, "null_prob": 0.05},
    "hours-per-week": {"type": "num", "range": (0, 80), "integer": True},
    "sex":            {"type": "cat", "values": ["Male", "Female", "Other"],
                       "weights": [0.5, 0.45, 0.05]},
    "income":         {"type": "cat", "values": ["<=50K", ">50K"],
                       "weights": [0.76, 0.24]},
}

# Optional: persist/share the schema
save_schema(schema, "adult_schema.json")          # or .yaml
schema = load_schema("adult_schema.json")

model = GReaT(llm="tabularisai/Qwen3-0.3B-distil")
mock = model.mock(
    schema=schema,
    n_samples=100,
    conditions={"age": ">= 40", "sex": "!= 'Male'"},   # same syntax as sample()
    examples=[                                          # few-shot rows for realism
        {"age": 35, "sex": "Female", "income": "<=50K"},
        {"age": 52, "sex": "Male",   "income": ">50K"},
    ],
    seed=42,                                            # reproducible across runs
)

Schema features:

• Per-column types: "num" with range, plus optional integer, precision, dist="normal" + mean + std. "cat" with values and optional weights (list or dict).
• null_prob (0–1) injects realistic missing values per column after generation.
• conditions= uses the same operator vocabulary as sample() — they tighten the schema's implicit constraints.
• examples= prepends few-shot demonstration rows to the prompt — big realism boost for LLMs that don't natively know the GReaT row format.
• seed= makes the entire output deterministic.
• save_schema / load_schema support JSON and YAML for sharing schemas across teams or environments.

Efficient Fine-Tuning with LoRA

GReaT supports LoRA (Low-Rank Adaptation) for parameter-efficient fine-tuning. This drastically reduces memory usage and training time, making it possible to fine-tune larger models on consumer hardware.

pip install peft

# LoRA with auto-detected target modules (works across model architectures)
model = GReaT(
    llm='meta-llama/Llama-3.1-8B-Instruct',
    batch_size=32,
    epochs=5,
    efficient_finetuning="lora",
    fp16=True,
)
model.fit(data)
synthetic_data = model.sample(n_samples=100)

You can also customize the LoRA hyperparameters:

model = GReaT(
    llm='tabularisai/Qwen3-0.3B-distil',
    batch_size=32,
    epochs=5,
    efficient_finetuning="lora",
    lora_config={
        "r": 8,
        "lora_alpha": 16,
        "lora_dropout": 0.1,
        "target_modules": ["q_proj", "v_proj"],  # optional, auto-detected if omitted
    },
    fp16=True,
)
model.fit(data)

Live Quality Monitoring During Training

Pass a held-out evaluation set to fit() to track synthesis quality as the model trains — no need to wait for training to finish. After each epoch (or every N steps), GReaT samples from the in-training model, scores it against eval_data using ColumnShapes similarity, and shows the score live in the chugchug progress bar.

from sklearn.model_selection import train_test_split

train_df, eval_df = train_test_split(data, test_size=0.2, random_state=42)

model = GReaT(llm="distilgpt2", epochs=10, batch_size=32)
model.fit(
    train_df,
    eval_data=eval_df,      # held-out comparison set
    eval_n_samples=50,      # synth rows drawn per evaluation (default 50)
    eval_every=None,        # None = end of every epoch; int = every N steps
)

You'll see q=0.612 appear in the chugchug bar (alongside loss=…, lr=…, epoch=…) plus a printed line 🎯 [quality @ step N] column_shapes_mean = 0.XXXX after each evaluation. column_shapes_mean lives in [0, 1]; closer to 1.0 means the synthetic marginal distributions are indistinguishable from the real ones. Sampling during training is the cost driver — keep eval_n_samples small (20–50) for fast feedback, larger for a more stable score.

GReaT Metrics

GReaT ships with a built-in evaluation suite to measure the quality, utility, and privacy of your synthetic data. All metrics follow the same interface:

from be_great.metrics import ColumnShapes, DiscriminatorMetric, MLEfficiency, DistanceToClosestRecord

# real_data: original pd.DataFrame
# synthetic_data: generated pd.DataFrame from model.sample()

ColumnShapes().compute(real_data, synthetic_data)
DiscriminatorMetric().compute(real_data, synthetic_data)
MLEfficiency(model=RandomForestClassifier, metric=accuracy_score,
             model_params={"n_estimators": 100}).compute(
    real_data, synthetic_data, label_col="target"
)
DistanceToClosestRecord().compute(real_data, synthetic_data)

Statistical Metrics

Metric	What it measures
ColumnShapes	Per-column distribution similarity (KS test for numerical, TVD for categorical)
ColumnPairTrends	Preservation of pairwise correlations (Pearson and Cramer's V)
BasicStatistics	Comparison of mean, std, and median per column

Fidelity & Utility Metrics

Metric	What it measures
DiscriminatorMetric	Trains a classifier to distinguish real from synthetic — score near 0.5 is best
MLEfficiency	Trains on synthetic, tests on real — measures downstream task utility

Privacy Metrics

Metric	What it measures
DistanceToClosestRecord	Distance from each synthetic record to its nearest real neighbor
kAnonymization	Minimum equivalence class size (higher = better privacy)
lDiversity	Diversity of sensitive attribute values within groups
IdentifiabilityScore	Risk of linking a synthetic record back to a specific real individual
DeltaPresence	Fraction of real records that have a near-exact synthetic match
MembershipInference	Simulated attack: can an adversary detect training set members?

GReaT Citation

If you use GReaT, please link or cite our work:

@inproceedings{borisov2023language,
  title={Language Models are Realistic Tabular Data Generators},
  author={Vadim Borisov and Kathrin Sessler and Tobias Leemann and Martin Pawelczyk and Gjergji Kasneci},
  booktitle={The Eleventh International Conference on Learning Representations },
  year={2023},
  url={https://openreview.net/forum?id=cEygmQNOeI}
}

Custom Synthetic Data

Need synthetic data for your business? We can help! Contact us at [e-mail] for custom data generation services.