> 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.

# Training configuration

A training job uses a dataset, workspace, training configuration and training compute configuration to produce a model. Prepare these settings before running training, then inspect the model, configure deployment and evaluate the deployed model. Start with an [available dataset](/ui-guides/data/datasets) and a [workspace](/ui-guides/models/workspaces). Complete the deployment steps before evaluating a newly trained model.

Use the example forms and resource records to locate the settings for your own run. Check names, dataset references, machine sizes and runtime limits before submitting a job. The training, evaluation and deployment examples use different resources; confirm the selected model at each stage.

## Before you begin

Decide which prediction task and dataset the configuration will support. Most projects can start with the prefilled induction settings. Change them when you need to control sampling, model structure, training duration, regularisation or resource use.

Prepare a separate [training compute configuration](/ui-guides/models/training-compute) for the machine. Saving either configuration does not start training.

## Create a training configuration

1. Open **Training configurations** and select **Create**.

2. Enter a **Configuration name**.

3. Review the settings throughout the form. Keep the defaults unless your modelling task requires a specific change.

4. Select **Create** to save the configuration. Open the saved record to check its settings before using it in a job.

The form groups settings into **Learning & optimization**, **Regularization**, **Tree & boosting**, **Binning**, **Interactions & complexity**, **Sampling & data**, **Partitioning & pruning**, and **Model capabilities**. These groups control optimisation, model structure and data handling. The controls below follow the same groups as the form. The starting values are the values prefilled by the UI; inspect a saved configuration for the values used by a particular job.

![Training configuration showing optimization, regularization and tree settings.](/_fern-files/_img/4dd7127c82978b2a362fcd84462c58aca77130767bbfbdf45d32ea1a40400d3b.webp)

## Review the configuration

Scroll through the configuration before saving. The example uses **DTree** binning with **32** bins, **64** maximum interactions, interaction degree **3**, polynomial degree **2**, and training and validation sizes of **0.8** and **0.2**. Check these choices for your own run. The training-validation split does not replace a separate test population for evaluation.

![Binning, interaction limits and sampling settings in the training configuration.](/_fern-files/_img/0fa1561905d3c1b3a448eaef5ccccdbccb24cf3851553ed984fae47a521399dd.webp)

![Partitioning, pruning and model capability controls at the bottom of the form.](/_fern-files/_img/577a23689c1a802715c283d052012755eb76b326292a5c916d3ae0c44734fd05.webp)

## Learning and optimisation

Learning-rate settings control how quickly training updates the model and what happens when progress stalls. Iteration and round limits bound the training loop; a job's maximum runtime provides a separate time limit.

| Control                 | UI starting value | Purpose                                                             |
| ----------------------- | ----------------- | ------------------------------------------------------------------- |
| Learning rate           | 1                 | Initial learning rate for the model.                                |
| Learning rate decay     | 1                 | Multiplier applied to the learning rate when loss does not improve. |
| Learning rate patience  | 16                | Iterations without improvement before learning-rate decay.          |
| Early stopping patience | 10                | Iterations without significant improvement before early stopping.   |
| Max iterations          | 10000             | Maximum iterations for each estimator.                              |
| Max rounds              | 50                | Maximum training rounds unless early stopping occurs.               |

If a run stops before learning useful structure, review the stopping settings and the job status together. Increasing limits can increase runtime; it does not guarantee a better model.

## Regularisation

Regularisation controls how strongly the model penalises coefficients. Use it when investigating overfitting or unnecessary model complexity, and compare the resulting model on the same evaluation population.

| Control        | UI starting value | Purpose                                                                  |
| -------------- | ----------------- | ------------------------------------------------------------------------ |
| Alpha          | 0                 | Elastic net regularisation strength.                                     |
| Alpha factor   | 2                 | Factor used for log-spaced search around Alpha.                          |
| L1 ratio       | 0                 | Balance between L1 and L2 regularisation.                                |
| L1 ratio delta | 0.2               | Search range around L1 ratio.                                            |
| Sparse model   | On                | Allows training to remove low-impact modules using the pruning settings. |

## Tree and boosting

These controls govern partitions, estimators and the order in which modules are boosted.

| Control                    | UI starting value            | Purpose                                                                                                                                   |
| -------------------------- | ---------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------- |
| Boosting strategy          | Interaction degree ascending | Choose **No grouping**, **Interaction degree ascending** or **Interaction degree descending** to control module grouping during boosting. |
| Estimators                 | 10                           | Number of estimators to train. More estimators can increase runtime, resource use and storage.                                            |
| Estimators winsorize level | 0                            | Proportion winsorized at each tail when aggregating estimator outputs.                                                                    |
| Initial depth              | 0                            | Initial partition depth.                                                                                                                  |
| Max depth                  | 6                            | Maximum partition depth.                                                                                                                  |
| Max depth momentum         | 1                            | Increment used to increase maximum depth during training. Enter a whole number.                                                           |

## Binning

Binning divides feature values into regions used in symbolic conditions.

| Control          | UI starting value | Purpose                                                                                       |
| ---------------- | ----------------- | --------------------------------------------------------------------------------------------- |
| Binning strategy | DTree             | Choose **Uniform**, **Quantile**, **DTree** or **GBTree** for the method used to create bins. |
| Number of bins   | 32                | Maximum bins per feature.                                                                     |
| Bin min distance | 0                 | Minimum distance between split points.                                                        |

## Interactions and complexity

Interaction limits determine how features may be combined. Complexity settings control how long training remains at a given complexity before attempting to increase it.

| Control                   | UI starting value | Purpose                                                               |
| ------------------------- | ----------------- | --------------------------------------------------------------------- |
| Max interactions          | 64                | Maximum interaction modules for each interaction degree.              |
| Max interaction degree    | 3                 | Maximum features combined in one interaction module.                  |
| Max polynomial degree     | 2                 | Highest feature degree in polynomial expansions.                      |
| Complexity max iterations | 1                 | Maximum consecutive training iterations at the same model complexity. |
| Complexity patience       | 1                 | Iterations without improvement allowed after increasing complexity.   |

Higher interaction degrees and polynomial degrees allow more complex relationships, but may require more resources. Change them deliberately and inspect both performance and the learned structure.

## Sampling and data

| Control         | UI starting value | Purpose                                                                                                                         |
| --------------- | ----------------- | ------------------------------------------------------------------------------------------------------------------------------- |
| Max samples     | 1                 | Fraction of the dataset used for training and validation. A value of 1 uses the full eligible population.                       |
| Subsample       | 10                | Fraction or number of training shards used to fit modules during boosting. Interpret this together with the dataset's sharding. |
| Training size   | 0.8               | Proportion of sampled data reserved for training.                                                                               |
| Validation size | 0.2               | Proportion reserved for validation.                                                                                             |

Use a fraction of at most 1 for **Max samples**. Review the training and validation proportions together. Reserve independent test observations for final evaluation; changing this split does not create a test population.

## Partitioning and pruning

Partition settings control when a split is worthwhile. Pruning settings control removal of low-impact modules when sparse modelling is enabled.

| Control                      | UI starting value | Purpose                                                                         |
| ---------------------------- | ----------------- | ------------------------------------------------------------------------------- |
| Partition min delta          | 0                 | Minimum absolute loss improvement required for a split.                         |
| Partition min delta factor   | 0.5               | Multiplier applied to the fractional threshold per module depth.                |
| Partition min delta fraction | 0.4               | Minimum fractional loss improvement required for a split.                       |
| Partition min samples        | 20                | Minimum training samples in a partition.                                        |
| Prune patience               | 1                 | Boosting rounds a module can remain below the pruning threshold before removal. |
| Prune threshold              | 0                 | Threshold below which ElasticNet coefficients are pruned.                       |

Prefer **Partition min delta fraction** unless you know the absolute loss improvement to require. Review pruning together with **Sparse model** rather than changing either setting in isolation.

## Model capabilities

| Control          | UI starting value | Purpose                                                            |
| ---------------- | ----------------- | ------------------------------------------------------------------ |
| Log interval     | 10                | Frequency of training progress logging.                            |
| Has causal       | Off               | Requests causal capabilities for a workflow that supports them.    |
| Estimators data  | Off               | Requests estimator capabilities for a workflow that supports them. |
| Exclude implicit | Off               | Restricts modules to those defined by model constraints.           |

> **Warning**
>
> Leave **Has causal** and **Estimators data** off unless your workflow supports them. Enabling a capability does not establish that a learned relationship is causal.
>
> **Exclude implicit** must not be combined with **Max interactions**. The UI form supplies Max interactions, so leave Exclude implicit off in this workflow. Use the [constraint configuration guidance](/guides/models/training/constraints) for a configuration that explicitly defines allowed modules.

## Reuse a configuration

Open the saved configuration before selecting it in another job. Check its name and settings, then use it with the intended dataset, workspace and training compute configuration. For a controlled comparison, create a clearly named configuration for the changed settings and retain each job's configuration reference.

## Next steps

Continue to [Training jobs](/ui-guides/models/training-jobs).