Sifar Labs builds datasets to spec for machine learning teams. You set the schema, the volume and the cases that matter. We source, generate and clean the data, then prove it holds up with a full fidelity report.
Tabular data, text and documents, for fraud, risk, insurance, health and legal AI teams. Based in Toronto.
Sample build: retail panel, 11 features
avg_monthly_spend_cad
98.3% match
Most teams hit the same wall. The architecture works and the infrastructure is ready, but the data is sparse, skewed, stuck behind a privacy review, or missing the cases that matter most. Sourcing and cleaning real data takes weeks. Building it from scratch takes longer.
When the cases that matter are a few hundred rows against millions, we generate realistic examples at whatever share you ask for, so your rules and models have enough to learn from and be tested against.
Fraud, chargebacks, defaults, severe claims
Records that behave like your customers, patients or members, with no real person behind any of them. Your team can build, test and demo without waiting on an access request.
Banking, credit, health records, insurance
Benchmarks where the right answer is already known, with fraud planted, errors seeded or fact patterns verified. Measure exactly what your model catches and what it misses.
Model evaluation, rule backtesting, vendor trials
A new loan, a new insurance line, a new client to onboard. We build the history your models need before the real one exists.
Product launches, new markets, cold starts
Formats: tabular data, text and documents, delivered as CSV, Parquet or JSON to your schema.
One short call to scope it, one week to a free sample. If the sample holds up, the full dataset follows the same four steps.
A 15-minute call covers your use case, schema, row count and the edge cases that matter.
We build the foundation from your own records or from public and licensed sources, keeping only what reflects your problem.
Generative models expand the data to your volume, with the rare classes and feature coverage you asked for.
A clean, structured dataset in your format, with a fidelity report that shows every metric before it touches your pipeline.
This is what separates a dataset from a CSV with a promise attached. Every delivery ships with a full statistical audit of how closely the generated data matches the source, feature by feature.
How closely the generated data reproduces the source, measured at three levels. Univariate checks each feature on its own. Bivariate checks every pair. Trivariate checks three-way interactions, where most synthetic data starts to slip.
98.7%
Average univariate accuracy
| Feature | Univariate | Bivariate | Trivariate |
|---|---|---|---|
province | 99.7% | 97.9% | 96.0% |
urban_rural_classification | 99.6% | 98.0% | 96.3% |
primary_store_format | 99.2% | 97.9% | 96.1% |
household_size | 99.2% | 98.0% | 96.3% |
age_band | 99.1% | 97.1% | 95.3% |
basket_size_avg | 98.8% | 97.7% | 96.4% |
visit_frequency_monthly | 98.8% | 97.3% | 95.6% |
avg_monthly_spend_cad | 98.3% | 97.2% | 95.5% |
churn_risk_label | 98.0% | 97.4% | 96.3% |
promotion_sensitivity | 97.9% | 96.9% | 95.4% |
loyalty_program_member | 97.4% | 97.0% | 95.9% |
| Total | 98.7% | 97.5% | 95.9% |
Data that gets each column right but breaks the relationships between them is useless for training. You get the full correlation matrix for the source and the generated data side by side, plus a difference matrix. When relationships survive, the difference matrix is nearly empty.
Every feature gets its own chart, with source and generated overlaid and labelled with its accuracy score, so you can see at a glance whether each column was faithfully reproduced.
98.7%
Average univariate accuracy
Feature interactions are where most generated datasets fall apart. We plot the joint distributions of the pairs that matter most, source beside generated, so you can check the relationships survived.
97.5%
Average bivariate accuracy
avg_monthly_spend_cad against visit_frequency_monthlychurn_risk_label against loyalty_program_memberpromotion_sensitivity against visit_frequency_monthlyHorizontal: visit_frequency_monthly. Vertical: avg_monthly_spend_cad.
Does the generated data sit in the same region of feature space as the source? Cosine similarity near 1.0 says it points the same way. A discriminator, a model trained to tell real records from generated ones, scores near 50% when it can't.
0.99880
Cosine similarity to source
Discriminator AUC
Two failure modes make generated data risky in production: copying the training records, and failing to generalize. Distance to closest record checks how near each generated row sits to a real one. A DCR share near 50% means generated rows are no closer to the training records than to records held out.
49.8%
DCR share
NNDR ratio 1.363, consistent with strong generalization
DCR share
Figures from a recent sample build on an 11-feature Canadian retail panel. Charts are illustrative.
For regulated teams, we send a validation script that runs inside your own environment against your real data. Only the summary numbers come back to us, so you can check our work without a single record leaving your systems.
Your environment
Your real data Our validation scriptYou can, and the open tools are good. The expensive part is your team's time: learning the tooling, modelling your schema, making the rare cases realistic and proving the result holds up.
| Option | Your own team | Self-serve software | Sifar Labs |
|---|---|---|---|
| Who does the work | Your engineers, around their other priorities | Your engineers, once they learn the tool | We do, start to finish |
| Time to a first dataset | Weeks | Days to weeks | Days, with a free sample inside a week |
| Proof of quality | Whatever checks you build | The tool's built-in reports | A full fidelity report, plus a script you can run on your own data |
| What you pay for | Salaried hours | An annual licence | A finished dataset, at a fixed fee |
Tell us your use case and we'll build a free 100-row sample to your spec within the week. No commitment. If it looks right, we scope the full dataset together.
Fifteen minutes on your model, your current data and what you're trying to build.
100 rows built to your spec, with its fidelity report, delivered within the week.
If the sample holds up, we agree on scope, timeline and delivery.
Your email app should have opened with everything filled in. Hit send and we'll reply within one business day. If nothing opened, write to us at Sales@SifarLabs.com.