Model Comparison Analysis (Model Compare)
Compare models registered in Model Registry side by side to review training configuration, metrics, and loss curves, and select the optimal model before assigning an alias.
Prerequisites
- Account with access to AI Training
- Model Registry must have 2 or more models in the same category
- At least one project exists
Step 1. Enter Model Compare
- Click Evaluation → Model Compare in the left sidebar.
- The Compare page opens with the following guidance at the top.
Select a model category and two models to compare their training configuration, metrics, and loss curves.
Model Compare targets only models registered in the Registry, and is used to review differences between versions before assigning an alias. You can proceed to Assign Alias in Model Registry based on the comparison results.
You can also reach this page via the Compare button at the top of the Model Registry list.
Step 2. Select Models to Compare
In the Section 1 area, select a category and 2 models to compare. You can compare exactly 2 models at a time.
Selection Fields
| Field | Input Method | Default | Description |
|---|---|---|---|
| Model category | Dropdown | Select category | Select from categories registered in Model Registry |
| Model (Option 1) | Dropdown | Select model | First comparison target from models in the selected category |
| Model (Option 2) | Dropdown | Select model | Second comparison target from models in the selected category |
Field Behavior
| Condition | Model (Option 1) / Model (Option 2) |
|---|---|
| Category not selected | Disabled, placeholder Select model |
| After category selected | Enabled, shows list of models in that category |
- Select the category containing the models you want to compare from the Model category dropdown.
- Select respective models in Model (Option 1) and Model (Option 2).
- Click the Run Comparison button.
Without Model Selection
If no category or models are selected, a guidance area is shown below.
- Select a model category
- Select options to compare their configurations, metrics, and training curves.
Step 3. View Model Details
After running the comparison, the Model details section displays metadata for both selected models in a 3-column table. The first column shows item names, the second and third columns show Option 1 and Option 2 values respectively.
INFO
| Item | Description |
|---|---|
| Experiment | Training experiment name |
| Base model | Base model used for training |
| Status | Model status |
| Created at | Model registration date and time |
PARAMETERS
| Item | Description |
|---|---|
| base_model | Base model identifier |
| batch_size | Training batch size |
| learning_rate | Learning rate |
| max_length | Maximum sequence length |
| max_steps | Maximum training steps |
| method | Training method |
| total_gpus | Number of GPUs used |
| training_mode | Training mode |
Review the differences in configuration values between the two models in each row, and narrow down the candidate version for alias assignment.
Step 4. View Metrics Comparison
Compare training result metrics in charts in the Metrics comparison section.
Loss Charts
| Metric | Axes |
|---|---|
| Training Loss | Y-axis: metric name / X-axis: loss value |
| Validation Loss | Y-axis: metric name / X-axis: loss value |
Compare the training and validation loss trends of the two models side by side to assess convergence speed and overfitting.
Runtime Chart
| Metric | Axes |
|---|---|
| Training Runtime (s) | Y-axis: metric name / X-axis: time in seconds (s) |
Compare training time to also evaluate resource efficiency.
Comparison Metric Summary
| Metric | Description |
|---|---|
| Training Loss | Loss on training data |
| Validation Loss | Loss on validation data |
| Training Runtime | Total training time (seconds) |
Step 5. Use Comparison Results (Assign Alias)
Once you've selected the superior model through comparison, assign an alias in Model Registry.
- Go to Evaluation → Model Registry.
- Click Assign Alias on the relevant category detail page.
- Assign an appropriate alias — Best, Active, or Custom — based on the comparison results.
| Alias | Description |
|---|---|
| Best | Best-performing version |
| Active | Currently in-use version |
| Custom | User-defined alias |