Skip to main content
After training completes, the Onyx Engine platform provides tools to analyze results, compare models, and download trained weights.

​
Jobs Tab

The Jobs tab shows all training and optimization jobs:
Jobs tab showing training history
Each job displays:
  • Status: Queued, Running, Completed, Failed
  • Model name: Target model for this job
  • Dataset: Source training data
  • Progress: Iterations completed
  • Metrics: Final loss values

​
Model Details

Click on a model to view detailed information:
Model details view

​
Overview

  • Configuration: Architecture, hyperparameters, features
  • Training history: Loss curves over iterations
  • Source dataset: Which data trained this model

​
Versions

Each training run creates a new model version. The platform tracks:
  • Version ID
  • Training date
  • Final metrics
  • Configuration differences

​
Test Predictions

Visualize model predictions on test data:
  • Ground truth: Actual values from dataset
  • Predictions: Model outputs
  • Error: Difference between predicted and actual

​
Lineage Tracking

View the complete data-to-model lineage:
Data lineage visualization
The lineage view shows:
  • Which raw datasets were processed
  • Which training datasets were created
  • Which models were trained from each dataset
  • Version history for each object

​
Comparing Models

To compare multiple models:
  1. Select models from the Table view
  2. Click “Compare”
  3. View side-by-side metrics and configurations
Comparison shows:
  • Architecture differences
  • Hyperparameter differences
  • Loss curve overlays
  • Final metric comparison

​
Downloading Models

​
Via Platform

  1. Navigate to the model
  2. Click “Download”
  3. Save the .pt file locally

​
Via SDK

from onyxengine import Onyx

# Initialize the client
onyx = Onyx()

# Download latest version
model = onyx.load_model('my_model')

# Download specific version
model = onyx.load_model('my_model', version_id='abc123...')

# Model is cached locally at ~/.onyx/models/

​
Understanding Metrics

​
Single-Step Loss

  • Measures one-step prediction accuracy
  • Lower is better
  • Good baseline metric for model quality

​
Multi-Step Loss

  • Measures trajectory simulation accuracy
  • More relevant for deployment
  • Sensitive to error accumulation

​
Interpreting Values

Loss RangeInterpretation
< 0.001Excellent fit
0.001 - 0.01Good fit
0.01 - 0.1Moderate fit, may need improvement
> 0.1Poor fit, check configuration
Loss values are relative to your data. Compare across models trained on the same dataset.

​
Optimization Results

For optimization jobs, additional information is available:

​
Trial Comparison

  • Each trial’s configuration
  • Metrics for each trial
  • Best trial identification

​
Hyperparameter Analysis

  • Which parameters had the biggest impact
  • Correlation between parameters and performance
  • Recommended configurations

​
Managing Versions

​
Setting Default Version

  1. Navigate to the model
  2. Go to Versions tab
  3. Click “Set as Default” on your preferred version
The default version is returned by load_model() without a version ID.

​
Deleting Versions

  1. Navigate to the model version
  2. Click “Delete”
  3. Confirm deletion
Deleted versions cannot be recovered. Downloaded copies remain valid.

​
Exporting Results

​
Export Metrics

Download training metrics as CSV:
  • Loss history
  • Learning rate schedule
  • Validation metrics

​
Export Configuration

Download model configuration as JSON:
  • Architecture parameters
  • Feature definitions
  • Training settings

​
Next Steps

Simulating Models

Deploy your trained model

Model Optimization

Find the best hyperparameters