Interested in trying Model Core? Contact your Invert team to request beta access. Features and labels may change before the general release.
Find your models
Open Models from the sidebar under Data management, or find it in the command palette.
The page has two tabs:
- Models: Saved models, including what each predicts, how many training runs it uses, and when it was trained.
- Specs: Reusable descriptions of modeling approaches you can use to build a model.
On the Models tab, filter by Personal to see models visible only to you, or Workspace to see models shared with your organization. Select All to see both. Turn on Archived only to find archived models.
Models with work in progress show an Active session badge. If your input is needed, the badge shows Needs input. Click the badge to open the associated Assist session.
Build a model
Open Assist and describe what you want to predict, or select the Train a model on my runs. starter. You can also start from a Model Spec to follow an established modeling approach.
Start from a Model Spec
Model Specs describe approaches such as scale-down model qualification or soft sensors. They are curated for your organization by Invert.
Open the Specs tab and search for a spec. Each card shows its model family and intent. Open a spec to review:
- Its approach, model family, intent, and process phase
- When to use it, when to avoid it, requirements, and known limitations
- Required input types and quantities, target output, and deliverables
- Development steps, validation checks, and reported metrics
Click Ask Assist to ask about the spec or start building a model from it. You can also attach a spec to an Assist chat as context.
Each model keeps a snapshot of the spec used to build it. Later changes to the spec do not change what the model reports. To find archived specs, turn on Archived only.
Define the target, inputs, and runs
Before training begins, you’ll confirm three things:
- Target: What the model predicts, including its units and an exact definition.
- Inputs: Which metrics, properties, and formulas the model can use. Exclude inputs that represent the target in another form or are only known after the fact.
- Runs: Which runs to use for training and which to hold back for validation and testing.
Train and save
Invert creates the model and its Development record, a private report containing the training data, code, and results. Headline performance metrics come from cross-validation or held-out runs, never from the runs used to fit the model.
Training can take a long time. You can leave the session and return later. The model keeps its Active session badge until the work is complete.
Once ready, the model is saved to Model Core. If you train the same model again, you’ll be asked whether to replace the saved model or create a new one.
Review a model
Open a model to see what it predicts, how it performs, and the evidence behind it.
Performance and supporting evidence
The Summary describes the result and includes up to four charts. Key metrics shows the reported performance metrics. Hover over a metric to read what it means.
For the supporting evidence, open Details > View record. The Development record contains the data, code, and results used to build and evaluate the model.
Inputs and training data
Target output shows what the model predicts. Required inputs shows the data a run must have before you can apply the model.
The Runs section links to the Training, Validation, and Test runs. Under Details, you can also find the training runs, the Model Spec used, and the original training session through Open in Assist.
Built on shows other models used to build this model. Used by shows models that depend on it.
Select Ask Assist to ask about the model, or Retrain in Assist to start a new training session based on it.
Apply a model to new runs
You can use saved models in Assist or in a report.
In Assist
Ask to apply a saved model to new runs. To attach a model to the chat, use Add > Models or an @ mention.
Before the first prediction, Invert checks that:
- The model reproduces the results in its Development record using the current data.
- Each run has all required inputs.
If a check fails, you’ll see an explanation before any predictions are made. Predictions appear in the chat and are not stored.
In a report
- Open Inputs.
- Click Add input and select the Model tab.
- Select a saved model and give it a name.
In Python code blocks, the model is available under that name. Its prediction function is available as <name>_predict.
The report keeps its own copy of the model. Select Refresh model to load a newer saved version, or Remove model to detach it. Removing a model from a report does not affect the saved model in Model Core.
Share and manage models
Share a model with your organization
New models are Personal and visible only to their author. After reviewing a model, the author can select Promote to workspace to make it available to everyone in the organization.
Before promoting, review the model’s summary, key metrics, labels, descriptions, and supporting evidence. Promotion confirms that you have reviewed and accepted them.
Promotion is permanent. The Development record becomes read-only for everyone and appears in your organization’s report list. To change a promoted model, train a new one or archive it.
Only trained models can be promoted. If a model depends on other Personal models, promote those first.
Archive a model
Select Archive to remove a model from the default list. Archived models remain available in Assist, so earlier comparisons continue to work.
Only the author can archive a Personal model. Anyone with write access can archive a Workspace model.
If a saved model uses a metric, formula, run, or material stream, that item cannot be archived in the Library. Invert lists the models blocking the action. Archive those models first, then archive the item.