> For clean Markdown of any page, append .md to the page URL.
> For a complete documentation index, see https://docs.umnai.com/llms.txt.
> For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.umnai.com/_mcp/server.

# Model Lab

Model Lab turns prepared datasets into explainable models. A training job learns from a dataset using saved training settings and compute resources. Its output is a trained model that you can inspect and deploy.

## How the resources fit together

| Resource                       | Role                                                                             |
| ------------------------------ | -------------------------------------------------------------------------------- |
| Dataset                        | The features and target prepared through onboarding.                             |
| Workspace                      | The context for training jobs, models and their metadata.                        |
| Training configuration         | The learning settings used to build the model.                                   |
| Model constraints              | Optional choices that guide which modules are included, excluded or prioritised. |
| Training compute configuration | The machine resources used to execute training.                                  |
| Training job                   | A run that brings these inputs together and produces a model.                    |
| Trained model                  | The saved training output, with references to its dataset, workspace and job.    |

A training configuration defines how the model learns. A compute configuration defines where the job runs. Keeping them separate lets you reuse settings and understand what changed between runs.

## From training to evaluation

Start with an available dataset and a workspace, then prepare the training settings and run a job. Inspect the output model and its resource references before deploying it. The [Training overview](/ui-guides/models/training-overview) explains this flow.

A trained model and a deployed model are distinct resources. **Serving** creates the running deployment used for queries. **Model Lab > Model Evaluation** evaluates that deployment, so deployment comes before evaluation even though evaluation is grouped in Model Lab.

## Inspect structure and behaviour

A model's learned structure includes modules, partitions and decision rules. Structure explains how the model is organised; attributions explain how its parts contributed to a particular prediction. In the app, use **Explanation Lab** to explore these views after deployment.

When comparing training runs, retain the dataset, configuration, constraints and job references. Compare evaluation results using the same data population and settings so that the differences are meaningful. See [Trained models](/ui-guides/models/models), [Evaluate models](/ui-guides/deployment/evaluation) and [Explanation views](/ui-guides/explanations/overview).

## Next steps

Create or choose a [workspace](/ui-guides/models/workspaces), then read [how training works](/ui-guides/models/training-overview).