Skip to main content
The Onyx Engine platform provides a visual interface for configuring and running training jobs.

​
Starting a Training Job

From a processed dataset, you can start training by:
  1. Navigate to your dataset in the System Overview
  2. Click “Train Model” from the dataset actions
  3. Configure the model and training parameters
  4. Start training

​
Feature Configuration

First, define which dataset columns are outputs vs inputs:
Feature configuration interface

​
Setting Up Features

1

Select Outputs

Choose which columns the model should predict (typically derivatives like acceleration)
2

Select Inputs

Choose columns to feed as model inputs (states and external inputs)
3

Define Relations

For state inputs, specify parent features and relation types (derivative, delta, equal)
4

Set Scaling

Choose normalization method for each feature

​
Example Configuration

For a system with acceleration, velocity, position, and control input:
FeatureTypeParentRelation
accelerationOutput--
velocityInputaccelerationderivative
positionInputvelocityderivative
control_inputInput--

​
Model Configuration

Choose your model architecture and hyperparameters:
Model configuration interface

​
Architecture Selection

Select from:
  • MLP: Multi-Layer Perceptron (fastest, good baseline)
  • RNN: Recurrent Neural Network (LSTM/GRU for temporal patterns)
  • Transformer: Attention-based (complex dependencies)

​
Common Parameters

ParameterDescriptionTypical Range
Sequence LengthInput history window1-16
Hidden LayersNetwork depth2-4
Hidden SizeNeurons per layer32-128
DropoutRegularization0.1-0.3

​
Training Parameters

ParameterDescriptionTypical Value
Training IterationsTotal training steps2000-5000
Batch SizeSamples per batch256-1024
Learning RateStep size3e-4
Checkpoint TypeOptimization targetsingle_step or multi_step

​
Running Training

Once configured:
  1. Click “Start Training”
  2. The job queues and starts when resources are available
  3. Monitor progress in the Jobs tab

​
Monitoring Training

View real-time training progress:
Training jobs view

​
Metrics Displayed

  • Training Loss: Error on training data (should decrease)
  • Validation Loss: Error on held-out data (watch for overfitting)
  • Learning Rate: Current LR if using a scheduler
  • Progress: Iterations completed / total

​
Understanding Loss Curves

  • Training loss steadily decreases
  • Validation loss follows training loss
  • Both converge to low values
  • Training loss decreases
  • Validation loss increases or plateaus
  • Fix: Increase dropout, reduce model size, add more data
  • Both losses remain high
  • Model not learning the dynamics
  • Fix: Increase model size, check feature configuration, verify data quality

​
Training Tips

Begin with an MLP and single_step checkpoint. Add complexity only if needed.
Check dataset statistics before training. Outliers or incorrect time steps cause poor results.
Start with 256-512. Larger batches train faster but may need learning rate adjustment.
If validation loss increases while training loss decreases, you’re overfitting.

​
Canceling Training

To stop a running job:
  1. Go to the Jobs tab
  2. Find your training job
  3. Click “Cancel”
Partial results may be saved depending on when you cancel.

​
Next Steps

Viewing Results

Analyze training results and download models

SDK Training

Train models programmatically