Skip to main content
An afternoon, most of it unattended. A Feature Refinery run on a small file takes tens of minutes to a few hours depending on how wide it is, and needs no supervision. Training is minutes. Prediction is minutes. Your own hands-on time is closer to ten.Skipping refinement and training directly takes minutes end to end.
Classification on labelled tabular data. Binary (fraud or not, good or bad risk) and multi-class.It is strongest where the answer that matters is rare: fraud, equipment failure, claims abuse, rare diagnoses. See Use cases and evidence.Feature Refinery handles binary targets only today. Multi-class data goes straight to Model Crucible.
Anything that is not labelled, since you need a known outcome per row to train from.The input has to be tabular, but that is a question of representation rather than subject matter. Images, signals and text all work once they are encoded as rows and columns: pixel or frequency features, embeddings, counts, whatever your domain already uses. If you can express it as a CSV with a target column, it is in scope.Regression can be approached by bucketing a continuous target into classes, but that is your modelling decision, not something the platform does for you.
A model scored on the rows it learned from can score them from memory, so the number means nothing. The 20% test file is held back and used only to mark the result. That held-back score is what you are shown and what the champion is chosen on.Keep the labels in both files. The test file needs its answers to mark with.
No. There are three settings, and the platform works out sensible values for all of them. The console is built for analysts.If you have data scientists, they go faster: the feature work that normally takes most of a project runs unattended, and they spend their time on the choice that actually matters, which candidate model fits the business trade-off.
The training target was set above what the data supports. Every attempt ran to its time cap and the run settled on a model that predicts one class for every row.Fix: use the dataset’s recommendedThreshold if it came from a refinement, or lower your target and run again. See Tuning.
Because a run averages many independent trainings and they do not all land in the same place. You can get a different baseline, different column decisions, and occasionally a run that removes nothing.Treat one run as evidence, not proof. If a particular column’s status matters to a decision you are making, run it twice.
Yes. MathFi.ai can run on your own cloud account or private cloud instead of on ours. That is arranged case by case rather than self-served.Everything in these docs describes the standard setup, where your tenant runs on our side and you call it at your own address. Talk to [email protected] before planning around anything else.
Datasets, refined files and trained models are stored under your account and are visible only to you. API access is by bearer token with a 60-minute idle window and a 24-hour absolute cap. Upload and download URLs are signed and expire in 60 minutes.For retention, residency or compliance questions, contact [email protected].
No. Training is a JVM process reading a CSV on ordinary CPUs. There is no accelerator to provision and no queue to wait in.
Yes. The same input gives the same decision every time. Every prediction records the model version it ran against, so changing the champion afterwards never changes what an earlier prediction meant, and any result can be replayed exactly.
CSV in, CSV out, over a REST API. That works with essentially any data tool.The API overview covers the mechanics, and Build a CLI or agent is a complete contract you can point a coding agent at to generate a client.There are no language SDKs today. The API is small enough not to need one.
Usage-based and agreed per organisation. Contact [email protected].
Yes, and you should. Train on the same split you use for your current model and compare on the same held-back rows.Published comparisons against tuned XGBoost, LightGBM, CatBoost, Random Forest and Logistic Regression across several public datasets are on the benchmarks page.
Still stuck? Email [email protected].