For AI / ML teams
Australian training and test data. No PII, no privacy review.
Privacy-preserving synthetic Australian demographic data, calibrated to ABS Census 2021, for ML training, customer profiling, demo data, and fairness testing.
Built for ML / AI workflows
- Hugging Face dataset
- Pandas / Polars / PyTorch / TensorFlow ready
- Parquet / CSV / Arrow
- Validated to SRMSE 0.05
- 8 demographic profiles
- Family + dwelling structure preserved
- 15,343 suburb granularity
- Geographically coherent
- Commercial license
Products
Self-serve. Pay as you go. Free tiers available.
Profile Scoring API
Send only the distinct, de-identified demographic combinations in any dataset (never raw records), and get the matching Australian profile for each. For segmentation, personas, and sizing.
1 credit per 100 combinations.
Individual-level Data
Bulk synthetic Australian population records (up to 27.5M), individual-level, ABS Census 2021 calibrated. For ML training, fine-tuning, demos.
Free 5,000-row sample on Hugging Face.
Validator (web + library)
Test an ML model for bias against Australian demographics, in the browser wizard or the open-source Python library. Free tier uses the 5k sample.
Free · MIT licensed.
Validator Pro
The full fairness audit against the entire national dataset (~27.5M records). Production-grade power; runs async in minutes.
50 credits per run.
Methodology, briefly
Generated from ABS Census 2021 conditional tables via Bayesian reconstruction with Gibbs sampling, preserving the real cross-tabs between age, sex, income, occupation, education, industry, household and dwelling.
Population marginals match real Census at 99.9%+; cross-tab fidelity sits at median person-level SRMSE 0.05 across 15,343 suburbs, adjusted to 2025–26. Read the full methodology →
Pricing
Free tier: 5 credits per week (never expires), the free 5k Hugging Face sample, and the open-source validator free tier.
Paid credits start at $10. 1 credit = 500 synthetic records; bundles never expire. See all pricing →
FAQ
Is synthetic data really good enough for ML training?
For population-level features (income, age, occupation, education, family structure) calibrated to a real Census, yes. AUSynth preserves the joint distributions between variables, not just the marginals, so models learn the same correlations they would from the underlying real data. The catch: synthetic data can't substitute for your real customer signal. Use AUSynth for the demographic substrate, your own data for the business outcomes you're modelling.
How does this compare to differential privacy?
Differential privacy adds calibrated noise to query results from a real dataset. AUSynth doesn't query a real dataset at all. Every record is generated from public Census conditionals, so there's no per-individual privacy budget to spend.
Can I use this for commercial models?
Yes. The credit bundles include commercial use rights. Cite as: Verosynthea AUSynth v1.0 (2026). verosynthea.com.
Can I use AUSynth to fine-tune LLMs?
Yes. We provide a row-to-text workflow that converts AUSynth records into natural-language descriptions suitable for LoRA fine-tuning, RLHF datasets, and instruction tuning. See /use-cases/llm-fine-tuning for the full pattern.
What's the difference between the Validator free and Pro tiers?
The free tier (web wizard or open-source Python library) tests your model against a 5,000-row sample, enough to catch obvious bias. Pro runs the same fairness metrics against the FULL national dataset (~27.5M records) for compliance-grade audits; runs are async (5–15 min) at 50 credits each.
Ready to try it?
Start with the free Hugging Face sample (no signup needed), or test a model for bias right now.