Synthetic Training Data
500K labeled edge cases.
None from real records.
Twynvex generates structurally faithful synthetic training data that fills the tail of your distribution — privacy-clean, schema-matched, and ready to merge with your pipeline.
The Coverage Problem
Real datasets have a coverage problem.
Your model fails on rare events. Not because you wrote bad code — because rare events weren't in your training set.
Distribution tails exist in every real-world dataset. Fraud at 0.1% frequency. Rare medical diagnoses. Unusual transaction patterns. Your model never learned them — it couldn't. Twynvex generates the examples your data collection never captured.
How It Works
From schema to synthetic rows in three steps.
No data sharing required. No manual annotation. Connect your schema and start generating.
Upload your schema + samples
Connect your existing dataset or define a schema via JSON. Twynvex auto-detects column types, distributions, and constraints.
import twynvex as twx
schema = twx.from_csv("transactions.csv")
# Auto-detected: 12 cols, 40K rowsConfigure generation targets
Specify edge case distribution, minority class ratios, and constraint rules. Twynvex fits marginal distributions and pairwise correlations.
job = schema.generate(
num_rows=500_000,
tail_weight=0.4,
label_ratio={"fraud": 0.15}
)Pull labeled synthetic rows
Download CSV, push to S3, or hit the REST API. Output schema is identical to source. Merge directly into your training pipeline.
df = job.to_dataframe()
# 500K rows, schema-matched
df.to_parquet("train_augmented.parquet")Capabilities
Schema-aware generation. Privacy by construction.
Every output row respects your referential integrity, your class constraints, and your pipeline format. No post-hoc filtering required.
Distribution-matched generation
Maintains joint distribution of real data with configurable shift for tail oversampling. JS divergence score reported on every generation run.
Schema-aware constraints
Referential integrity, categorical cardinality, and numeric bounds enforced at generation time. No post-hoc filtering required.
Privacy by construction
Zero real records in output. k-anonymity configurable. GDPR-compatible by design — no personal data ever processed by the generation engine.
Drop-in pipeline format
Output as CSV, Parquet, JSON Lines, or via REST API. Schema identical to source — no transformation step between Twynvex and your training loop.
Use Cases
Where real data runs out.
Three patterns where synthetic generation solves what more data collection cannot.
Class imbalance at 0.1% positive rate
SMOTE degrades below 0.5% class ratio. Twynvex generates schema-faithful synthetic fraud patterns from behavioral schema — no real transaction records needed.
Measurable AUC improvement on real held-out test set
Long-tail intents with <50 examples
Chatbot classifiers with 200+ intents fail on rare classes. Synthetic text samples generated for under-represented intents using structural templates and semantic variation.
JSONL output compatible with OpenAI fine-tuning and HuggingFace
Real patient records blocked by HIPAA
Model development stalls when data can't leave the hospital. Synthetic patient cohorts with configurable demographics, co-occurrence rates, and lab ranges — privacy by construction.
Privacy-clean synthetic cohort for cloud ML platform development
From the field
ML engineers on Twynvex.
Twynvex filled our fraud class from 0.08% to 15% in one generation run. The schema fidelity was close enough to real data that our downstream XGBoost model trained without any additional preprocessing changes.
We had a patient cohort that couldn't leave our on-prem environment. Twynvex let us generate a synthetic equivalent with matching demographic distributions and ship it to our cloud training cluster. The utility score came back at 0.91 — usable.
Your model is only as good as its training distribution.
Generate edge cases in minutes. No data sharing required.