Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.mathfi.ai/llms.txt

Use this file to discover all available pages before exploring further.

This guide outlines the core process of preparing data, train models and run predictions within MathFi.ai platform.

​
Overview

You can interact with MathFi.ai in two ways:
  1. Web Dashboard — For research, PoC, and interactive model training with hyperparameter tuning (what this guide is about)
  2. REST API/Python — For programmatic, production-ready use of your trained models via the MathFi.ai API (see API guide)
All data and models are securely logged and stored on the cloud under your account.
How classification type is determined
  • The MathFi.ai platform automatically detects whether the task is binary or multi-class classification.
  • A balanced test set is generated using your labelled training data.
  • High accuracy generally implies good F1 Score and AUC (Area Under the Curve).

​
Video walkthrough

Now you can continue reading this detailed step by step guide or watch this video walkthrough (present also in the Getting started guide) on how to create datasets, training models and predict outcomes within MathFi.ai platform.


​
Dataset creation


Before proceeding, ensure your dataset CSV is formatted correctly by following CSV Format Guide

  1. Login to the MathFi.ai dashboard.
  2. Click “Datasets” in the top-right menu. Step 1 - Datasets Menu
  3. Click ”+ Create” to start a new dataset. Step 2 - Create Dataset
  4. Enter a name for your dataset.
  5. Set the number of buckets: default is 20 (range: 4–100).
  6. Upload your labelled CSV file.
  7. Click “Save”. Step 3 - Save Dataset

After this, the dataset will start processing. This may take up to few minutes or hours depending on the size and complexity of the data.

​
Model training


The dataset must have finished processing before attempting to train a model with it

  1. Navigate to “Trainings” in the sidebar. Step 4 - Open Trainings
  2. Click “Create”. Step 5 - Create Training
  3. Configure the training:
    • Scaling Factor: (default: 19, range: 8–499)
    • Performance Threshold: (value between 0 and 1)
    • Select your dataset from the dropdown.

  4. Click “Save”. Step 6 - Save Training
The platform runs 4 algorithms in parallel and selects a new champion model if one outperforms the rest.

​
Understanding training results

  • The accuracy reflects the performance on a balanced, unseen 10% test set.
  • Logs and metrics are saved with the model.
  • Further training with tuned hyperparameters can improve performance. See Hyperparameter tuning guide

​
Prediction running


Predictions can be run once training has been completed and the first champion model is created

  1. Click “Predictions” in the left menu. Step 7 - Open Predictions
  2. Click “Create”. Step 8 - Create Prediction
  3. Select:
    • Your original training dataset
    • Your batch prediction CSV (see CSV Format Guide for instructions on how to create)

  1. Click “Save”. Step 9 - Save Prediction

​
Downloading results

Monitor progress on the Predictions listing page. Once finished, click “Download” to retrieve your prediction results. Step 10 - Download Predictions
Results include predicted labels and their probability scores. Step 10 - Download Predictions

​
Recap

StageDescription
DatasetUpload labelled CSV, set number of buckets
TrainingSet scaling factor & threshold, train model
PredictionUpload unlabelled CSV, get prediction results
See also: Hyperparameter tuning to improve model performance, and real-world examples in Healthcare and Credit Card Approval.

​
Regression use case

The MathFi.ai platform is mainly designed for classification. However, you can solve regression type prediction problems with high accuracy by converting continuous targets into discrete classes.

​
Example: Wind Turbine Power prediction

  1. Bin the output range (e.g., 0–2000 kW) into classes:
    • 0–200 kW → Class 1
    • 200–400 kW → Class 2
    • …
    • 1800–2000 kW → Class 10
  2. Train a multi-class classifier on these buckets.
  3. To increase granularity:
    • Take the winning bin (e.g., 400–600 kW)
    • Divide it further (e.g., 10 × 20 kW sub-bins)
    • Train again for finer predictions.

​
Troubleshooting NaN predictions

Predictions may occasionally return NaN probability values when the number of buckets is insufficient for the model to make a confident prediction. To resolve this, create a new dataset with a higher number of buckets and retrain. If NaN values persist, your training dataset may need more labelled samples.