likithreddy949/sdg-synthetic-data

SDG — Synthetic Data Generator

A support-aware, rule-governed, autoregressive synthetic data generator for tabular data.

What it does

Upload a CSV. The pipeline automatically:

• —Detects one-hot encoded column groups and collapses them
• —Identifies deterministic fields that can be computed from others
• —Records value bounds and integer constraints
• —Maps valid categorical combinations
• —Infers optimal generation order
• —Trains a GRU autoregressive model on stochastic fields only
• —Generates new synthetic rows with 100% structural validity
• —Returns a downloadable CSV of synthetic data

Benchmark results

Compared to CTGAN (SDV library) on identical splits:

Architecture

Six-piece pipeline: Rule System, Dependency Graph, Probabilistic Generator (GRU AR), Deterministic Resolver, Legality Shell, Evaluation Layer.

Philosophy: Learn only what is uncertain. Compute what is deterministic. Reject what violates support.

Tech stack

PyTorch · NumPy · Pandas · scikit-learn · Gradio