# Create custom models

> Create custom models - How to create and assemble a custom model in the workshop.

This Markdown file sits beside the HTML page at the same path (with a `.md` suffix). It summarizes the topic and lists links for tools and LLM context.

Companion generated at `2026-05-06T18:17:10.047640+00:00` (UTC).

## Primary page

- [Create custom models](https://docs.datarobot.com/en/docs/workbench/nxt-registry/nxt-model-workshop/nxt-create-custom-model.html): Full documentation for this topic (HTML).

## Sections on this page

- [Create a new custom model](https://docs.datarobot.com/en/docs/workbench/nxt-registry/nxt-model-workshop/nxt-create-custom-model.html#create-a-new-custom-model): In-page section heading.
- [Configure time series settings](https://docs.datarobot.com/en/docs/workbench/nxt-registry/nxt-model-workshop/nxt-create-custom-model.html#configure-time-series-settings): In-page section heading.
- [Assemble the custom model](https://docs.datarobot.com/en/docs/workbench/nxt-registry/nxt-model-workshop/nxt-create-custom-model.html#assemble-the-custom-model): In-page section heading.
- [Connect to remote repositories](https://docs.datarobot.com/en/docs/workbench/nxt-registry/nxt-model-workshop/nxt-create-custom-model.html#connect-to-remote-repositories): In-page section heading.
- [Drop-in environments](https://docs.datarobot.com/en/docs/workbench/nxt-registry/nxt-model-workshop/nxt-create-custom-model.html#drop-in-environments): In-page section heading.
- [Configure evaluation and moderation](https://docs.datarobot.com/en/docs/workbench/nxt-registry/nxt-model-workshop/nxt-create-custom-model.html#configure-evaluation-and-moderation): In-page section heading.
- [Define runtime parameters](https://docs.datarobot.com/en/docs/workbench/nxt-registry/nxt-model-workshop/nxt-create-custom-model.html#define-runtime-parameters): In-page section heading.
- [Add parameters](https://docs.datarobot.com/en/docs/workbench/nxt-registry/nxt-model-workshop/nxt-create-custom-model.html#add-parameters): In-page section heading.
- [Edit parameters](https://docs.datarobot.com/en/docs/workbench/nxt-registry/nxt-model-workshop/nxt-create-custom-model.html#edit-parameters): In-page section heading.
- [Assign custom model datasets](https://docs.datarobot.com/en/docs/workbench/nxt-registry/nxt-model-workshop/nxt-create-custom-model.html#assign-custom-model-datasets): In-page section heading.
- [Disable column filtering for prediction requests](https://docs.datarobot.com/en/docs/workbench/nxt-registry/nxt-model-workshop/nxt-create-custom-model.html#disable-column-filtering-for-prediction-requests): In-page section heading.
- [Configure custom model resource settings](https://docs.datarobot.com/en/docs/workbench/nxt-registry/nxt-model-workshop/nxt-create-custom-model.html#configure-custom-model-resource-settings): In-page section heading.
- [Select a resource bundle](https://docs.datarobot.com/en/docs/workbench/nxt-registry/nxt-model-workshop/nxt-create-custom-model.html#select-a-resource-bundle): In-page section heading.
- [Assemble an anomaly detection model](https://docs.datarobot.com/en/docs/workbench/nxt-registry/nxt-model-workshop/nxt-create-custom-model.html#assemble-an-anomaly-detection-model): In-page section heading.
- [Assemble a text generation model](https://docs.datarobot.com/en/docs/workbench/nxt-registry/nxt-model-workshop/nxt-create-custom-model.html#assemble-a-text-generation-model): In-page section heading.
- [Assemble a vector database](https://docs.datarobot.com/en/docs/workbench/nxt-registry/nxt-model-workshop/nxt-create-custom-model.html#assemble-a-vector-database): In-page section heading.
- [Assemble an agentic workflow](https://docs.datarobot.com/en/docs/workbench/nxt-registry/nxt-model-workshop/nxt-create-custom-model.html#assemble-an-agentic-workflow): In-page section heading.
- [Register an agentic tool](https://docs.datarobot.com/en/docs/workbench/nxt-registry/nxt-model-workshop/nxt-create-custom-model.html#register-an-agentic-tool): In-page section heading.

## Related documentation

- [NextGen UI documentation](https://docs.datarobot.com/en/docs/workbench/index.html): Linked from this page.
- [Registry](https://docs.datarobot.com/en/docs/workbench/nxt-registry/index.html): Linked from this page.
- [Workshop](https://docs.datarobot.com/en/docs/workbench/nxt-registry/nxt-model-workshop/index.html): Linked from this page.
- [custom](https://docs.datarobot.com/en/docs/classic-ui/mlops/deployment/custom-models/custom-model-environments/custom-environments.html#create-a-custom-environment): Linked from this page.
- [assembling a custom model](https://docs.datarobot.com/en/docs/api/code-first-tools/drum/index.html): Linked from this page.
- [test the contents locally](https://docs.datarobot.com/en/docs/api/code-first-tools/drum/custom-local-test.html): Linked from this page.
- [testing suite](https://docs.datarobot.com/en/docs/workbench/nxt-registry/nxt-model-workshop/nxt-test-custom-model.html): Linked from this page.
- [Proxyfor an external model](https://docs.datarobot.com/en/docs/classic-ui/mlops/deployment/custom-models/custom-model-workshop/ext-model-proxy.html): Linked from this page.
- [Binary](https://docs.datarobot.com/en/docs/reference/glossary/index.html#classification): Linked from this page.
- [Unstructured](https://docs.datarobot.com/en/docs/api/code-first-tools/drum/unstructured-custom-models.html): Linked from this page.
- [Location](https://docs.datarobot.com/en/docs/workbench/nxt-console/nxt-monitoring/nxt-data-drift.html#enable-geospatial-monitoring-for-a-deployment): Linked from this page.
- [model-metadata.yaml](https://docs.datarobot.com/en/docs/api/code-first-tools/drum/custom-model-metadata.html#inference-model-metadata): Linked from this page.
- [multiseries models](https://docs.datarobot.com/en/docs/classic-ui/modeling/time/multiseries.html): Linked from this page.
- [real-time predictions](https://docs.datarobot.com/en/docs/workbench/nxt-console/nxt-predictions/nxt-pred-api-snippets.html#real-time-prediction-snippet-settings): Linked from this page.
- [challengers](https://docs.datarobot.com/en/docs/workbench/nxt-console/nxt-mitigation/nxt-challengers.html): Linked from this page.
- [runtimeParameterDefinitions](https://docs.datarobot.com/en/docs/api/code-first-tools/drum/custom-model-runtime-parameters.html): Linked from this page.
- [register a new remote repository](https://docs.datarobot.com/en/docs/platform/acct-settings/remote-repos.html): Linked from this page.
- [Configure evaluation and moderation](https://docs.datarobot.com/en/docs/workbench/nxt-registry/nxt-model-workshop/nxt-configure-evaluation-moderation.html): Linked from this page.
- [register](https://docs.datarobot.com/en/docs/workbench/nxt-registry/nxt-model-directory/nxt-register-cus-models.html): Linked from this page.
- [deploy](https://docs.datarobot.com/en/docs/workbench/nxt-registry/nxt-model-directory/nxt-deploy-models.html): Linked from this page.
- [predictions](https://docs.datarobot.com/en/docs/workbench/nxt-console/nxt-predictions/index.html): Linked from this page.
- [Custom metrics](https://docs.datarobot.com/en/docs/workbench/nxt-console/nxt-monitoring/nxt-custom-metrics.html): Linked from this page.
- [Data exploration](https://docs.datarobot.com/en/docs/workbench/nxt-console/nxt-monitoring/nxt-data-exploration.html): Linked from this page.
- [data drift](https://docs.datarobot.com/en/docs/classic-ui/mlops/monitor/data-drift.html): Linked from this page.
- [accuracy](https://docs.datarobot.com/en/docs/classic-ui/mlops/monitor/deploy-accuracy.html): Linked from this page.
- [minor custom model version](https://docs.datarobot.com/en/docs/workbench/nxt-registry/nxt-model-workshop/nxt-custom-model-versions.html): Linked from this page.
- [the deprecated "per model" training data assignment method](https://docs.datarobot.com/en/docs/release/deprecations-and-migrations/cus-model-training-data.html): Linked from this page.
- [DRUM](https://docs.datarobot.com/en/docs/api/code-first-tools/drum/custom-model-drum.html): Linked from this page.
- [organization admins](https://docs.datarobot.com/en/docs/platform/admin/manage-entities/manage-users.html#set-admin-permissions-for-users): Linked from this page.
- [custom environment](https://docs.datarobot.com/en/docs/workbench/nxt-registry/nxt-environment-workshop/index.html): Linked from this page.
- [custom model code](https://docs.datarobot.com/en/docs/api/code-first-tools/drum/structured-custom-models.html): Linked from this page.
- [Bolt-on Governance API](https://docs.datarobot.com/en/docs/agentic-ai/genai-code/genai-chat-completion-api.html): Linked from this page.
- [libraries (and versions)](https://docs.datarobot.com/en/docs/classic-ui/mlops/deployment/custom-models/custom-model-workshop/custom-model-dependencies.html): Linked from this page.
- [Monitor workflows](https://docs.datarobot.com/en/docs/agentic-ai/agentic-monitor/index.html): Linked from this page.
- [View and manage tools](https://docs.datarobot.com/en/docs/workbench/nxt-registry/nxt-tools/nxt-view-manage-tools.html): Linked from this page.

## Documentation content

Custom models are user-created, pretrained models that you can upload to DataRobot (as a collection of files) via the Workshop. You can assemble custom models in either of the following ways:

- Create a custom modelwithoutthe environment requirements and astart_server.shfile on theAssembletab. This type of custom modelmustuse a drop-in environment. Drop-in environments contain the requirements andstart_server.shfile used by the model. They areprovided by DataRobotin the workshop.
- Create a custom modelwiththe environment requirements and astart_server.shfile on theAssembletab. This type of custom model can be paired with acustomordrop-inenvironment.

Be sure to review the guidelines for [assembling a custom model](https://docs.datarobot.com/en/docs/api/code-first-tools/drum/index.html) before proceeding. If any files overlap between the custom model and the environment folders, the model's files will take priority.

> [!NOTE] Testing custom models
> Once a custom model's file contents are assembled, you can [test the contents locally](https://docs.datarobot.com/en/docs/api/code-first-tools/drum/custom-local-test.html) for development purposes before uploading it to DataRobot. After you create a custom model in the workshop, you can run a [testing suite](https://docs.datarobot.com/en/docs/workbench/nxt-registry/nxt-model-workshop/nxt-test-custom-model.html) from the Test tab.

## Create a new custom model

To create a custom model in preparation for [assembly](https://docs.datarobot.com/en/docs/workbench/nxt-registry/nxt-model-workshop/nxt-create-custom-model.html#assemble-the-custom-model):

1. ClickRegistry > Workshop. This tab lists the models you have created:
2. Click one of the following options (or thebutton when the custom model panel is open), depending on the currently selected tab:
3. On theAdd a modelpage, enter the fields underConfigure the model: FieldDescriptionModel typeThe type of custom model to create, a standardCustom modelor aProxyfor an external model.Model nameA descriptive name for the custom model.Target typeThe type of prediction the model makes. Depending on the prediction type, you must configure additional settings:Binary: For a binary classification model, enter thePositive class labeland theNegative class label. Then, optionally, enter aPrediction threshold.Regression: No additional settings.Time Series (Binary):Preview feature. For a binary classification model, enter thePositive class labeland theNegative class labeland configure thetime series settings. Then, optionally, enter aPrediction threshold.Time Series (Regression):Preview featureConfigure thetime series settings.Multiclass: For a multiclass classification model, enter or upload (.csv,.txt) theTarget classesfor your target, one class per line. To ensure that the classes are applied correctly to your model's predictions, enter classes in the same order as your model's predicted class probabilities.Text Generation:Premium feature. No additional settings.Anomaly Detection: No additional settings.Unstructured: No additional settings. Unstructured models do not need to conform to a specific input/output schema and may use a different request format. Prediction drift, accuracy tracking, challengers, and humility are disabled for deployments with a "Unstructured" target type. Service health, deployment activity, and governance remain available.Vector Database:Preview featureNo additional settings.Location:Premium feature. No additional settings.Agentic Workflow:Premium feature. No additional settings.MCP Server:Premium feature. No additional settings. An MCP server is a tool that allows you to expose the DataRobot API as a tool that can be used by agents.Target nameThe dataset column the model predicts. Required for target types that use a target column, including multiclass (along withTarget classesfor class labels). For multiclass, the same column name belongs inmodel-metadata.yamlunderinferenceModel.targetName, and class names belong ininferenceModel.classLabels. This field is not used for anomaly detection models.Optional settingsPrediction thresholdFor binary classification models, a decimal value ranging from 0 to 1. Any prediction score exceeding this value is assigned to the positive class.LanguageThe programming language used to build the model.DescriptionA description of the model's contents and purpose. Target type support for proxy modelsIf you select theProxymodel type, you can't select theUnstructuredtarget type. PremiumVector database deployments are a premium feature, off by default and require GenAI experimentation. Contact your DataRobot representative or administrator for information on enabling this feature. PremiumAgentic workflows are a premium feature. Contact your DataRobot representative or administrator for information on enabling the feature.
4. After completing the fields, clickAdd model. The custom model opens to theAssembletab.

### Configure time series settings

> [!NOTE] Preview
> Time series custom models are off by default. Contact your DataRobot representative or administrator for information on enabling this feature.
> 
> Feature flag: Enable Time Series Custom Models, Enable Feature Filtering for Custom Model Predictions

You can create time series custom models by configuring the following time series-specific fields, in addition to the fields required for binary classification and regression models. To create a time series custom model, select Time Series (Binary) or a Time Series (Regression) as a Target type, and configure the following settings while creating the model:

| Field | Description |
| --- | --- |
| Date/time feature | The column in the training dataset that contains date/time values of the given prediction row. |
| Date/time format | The format of the values in both the date/time feature column and the forecast point feature column, provided as a dropdown list of all possible values in GNU C library format. |
| Forecast point feature | The column in the training dataset that contains the point from which you are making a prediction. |
| Forecast unit | The time unit (seconds, days, months, etc.) that the time step uses, provided as a dropdown list of all possible values. |
| Series ID feature | Optional. For multiseries models, the column in the training dataset that identifies which series each row belongs to. |

When you make [real-time predictions](https://docs.datarobot.com/en/docs/workbench/nxt-console/nxt-predictions/nxt-pred-api-snippets.html#real-time-prediction-snippet-settings) with a time series custom model, in the CSV serialization of the prediction response, any extra columns (beyond the prediction results) returned from the model have a column name suffixed with `_OUTPUT`.

> [!NOTE] Considerations for time series custom models
> Time series custom models:
> 
> Cannot be selected as
> challengers
> .
> Only support the model startup test during
> custom model testing
> .
> Support
> real-time predictions
> , not batch predictions.
> Do not support the portable prediction server (PPS).

## Assemble the custom model

After you [create a custom model](https://docs.datarobot.com/en/docs/workbench/nxt-registry/nxt-model-workshop/nxt-create-custom-model.html#create-a-new-custom-model), you can provide the required environment, dependencies, and files:

1. In the model you want to assemble, on theAssembletab, navigate to theEnvironmentsection and select a model environment from theBase environmentdropdown menu. Model environmentsThe model environment is used fortestingand deploying the custom model. TheBase Environmentdropdown list includesdrop-in model environmentsand anycustom environmentsthat you can create. By default, the custom model uses the most recent version of the selected environment with a successful build.
2. To populate theDependenciessection, you can upload arequirements.txtfile in theFilessection, allowing DataRobot to build the optimal image.
3. In theFilessection, add the required custom model files. If you aren't pairing the model with adrop-in environment, this includes the custom model environment requirements and astart_server.shfile. You can add files in several ways: ElementDescription1FilesDrag files into the group box for upload.2Choose from sourceClick to browse forLocal Filesor aLocal Folder. Topull files from a remote repository, use theUploadmenu.3UploadClick to browse forLocal Filesor aLocal Folderor to pull files from aremote repository.4CreateCreate a new file, empty or as a template, and save it to the custom model:Create model-metadata.yaml: Creates a startermodel-metadata.yamlfor this model’s target type (including requiredinferenceModelfields for binary and multiclass). You can addruntimeParameterDefinitionsin the same file.Create blank file: Creates an empty file. Click the edit icon () next toUntitledto provide a file name and extension, then add your custom contents. File replacementIf you add a new file with the same name as an existing file, when you clickSave, the old file is replaced in theFilessection. Model file locationIf you add files from aLocal Folder, make sure the model file is already at the root of the custom model. Uploaded folders are for dependent files and additional assets required by your model, not the model itself. Even if the model file is included in the folder, it will not be accessible to DataRobot unless the file exists at the root level. The root file can then point to the dependencies in the folder.

If you incorrectly add one or more files or folders to the Assemble tab, in the Files section, you can click the delete () icon next to each file or folder to remove it from the custom model.

### Connect to remote repositories

In the Files section, if you click Upload > Remote Repository, the Add the content from remote repository panel opens. You can click Remote Repositories to [register a new remote repository](https://docs.datarobot.com/en/docs/platform/acct-settings/remote-repos.html) (A), or select an existing remote repository (B):

After registering and selecting a repository, you can select the checkbox for each file or folder you want to upload, and then click Select content:

The selected files are pulled into the Files section.

### Drop-in environments

DataRobot provides drop-in environments in the workshop, defining the required libraries and providing a `start_server.sh` file. Each environment is prefaced with [DataRobot]in the Environment section of the Workshop's Assemble tab.

The available drop-in environments depend on your DataRobot installation; however, the table below lists commonly available public drop-in environments with [templates in the DRUM repository](https://github.com/datarobot/datarobot-user-models/tree/master/public_dropin_environments). Depending on your DataRobot installation, the Python version of these environments may vary, and additional non-public environments may be available for use.

**Managed AI Platform (SaaS):**
> [!NOTE] Drop-in environment security
> Starting with the March 2025 Managed AI Platform release, most general purpose DataRobot custom model drop-in environments are security-hardened container images. When you require a security-hardened environment for running custom jobs, only shell code following the [POSIX-shell standard](https://pubs.opengroup.org/onlinepubs/9799919799/utilities/sh.html) is supported. Security-hardened environments following the POSIX-shell standard support a limited set of shell utilities.

**Self-Managed AI Platform:**
> [!NOTE] Drop-in environment security
> Starting with the 11.0 Self-Managed AI Platform release, most general purpose DataRobot custom model drop-in environments are security-hardened container images. When you require a security-hardened environment for running custom jobs, only shell code following the [POSIX-shell standard](https://pubs.opengroup.org/onlinepubs/9799919799/utilities/sh.html) is supported. Security-hardened environments following the POSIX-shell standard support a limited set of shell utilities.


| Environment name & example | Compatibility & artifact file extension |
| --- | --- |
| Python 3.X | Python-based custom models and jobs. You are responsible for installing all required dependencies through the inclusion of a requirements.txt file in your model files. |
| Python 3.X GenAI Agents | Generative AI models (Text Generation or Vector Database target type) |
| Python 3.X ONNX Drop-In | ONNX models and jobs (.onnx) |
| Python 3.X PMML Drop-In | PMML models and jobs (.pmml) |
| Python 3.X PyTorch Drop-In | PyTorch models and jobs (.pth) |
| Python 3.X Scikit-Learn Drop-In | Scikit-Learn models and jobs (.pkl) |
| Python 3.X XGBoost Drop-In | Native XGBoost models and jobs (.pkl) |
| Python 3.X Keras Drop-In | Keras models and jobs backed by tensorflow (.h5) |
| Java Drop-In | DataRobot Scoring Code models (.jar) |
| R Drop-in Environment | R models trained using CARET (.rds) Due to the time required to install all libraries recommended by CARET, only model types that are also package names are installed (e.g., brnn, glmnet). Make a copy of this environment and modify the Dockerfile to install the additional, required packages. To decrease build times when you customize this environment, you can also remove unnecessary lines in the # Install caret models section, installing only what you need. Review the CARET documentation to check if your model's method matches its package name. (Log in to GitHub before clicking this link.) |

> [!NOTE] scikit-learn
> All Python environments contain scikit-learn to help with preprocessing (if necessary), but only scikit-learn can make predictions on `sklearn` models.

When using a drop-in environment, your custom model code can reference several environment variables injected to facilitate access to the [DataRobot Client](https://pypi.org/project/datarobot/) and [MLOps Connected Client](https://pypi.org/project/datarobot-mlops-connected-client/):

| Environment Variable | Description |
| --- | --- |
| MLOPS_DEPLOYMENT_ID | If a custom model is running in deployment mode (i.e., the custom model is deployed), the deployment ID is available. |
| DATAROBOT_ENDPOINT | If a custom model has public network access, the DataRobot endpoint URL is available. |
| DATAROBOT_API_TOKEN | If a custom model has public network access, your DataRobot API token is available. |

## Configure evaluation and moderation

> [!NOTE] Premium
> Evaluation and moderation guardrails are a premium feature. Contact your DataRobot representative or administrator for information on enabling this feature.
> 
> Feature flag: Enable Moderation Guardrails ( Premium), Enable Global Models in the Model Registry ( Premium), Enable Additional Custom Model Output in Prediction Responses

Evaluation and moderation guardrails help your organization block prompt injection and hateful, toxic, or inappropriate prompts and responses. It can also prevent hallucinations or low-confidence responses and, more generally, keep the model on topic. In addition, these guardrails can safeguard against the sharing of personally identifiable information (PII). Many evaluation and moderation guardrails connect a deployed text generation model (LLM) to a deployed guard model. These guard models make predictions on LLM prompts and responses and report these predictions and statistics to the central LLM deployment. To use evaluation and moderation guardrails, first, create and deploy guard models to make predictions on an LLM's prompts or responses; for example, a guard model could identify prompt injection or toxic responses. Then, when you create a custom model with the Text Generation target type, define one or more evaluation and moderation guardrails.

To select and configure evaluation and moderation guardrails, on the Assemble tab for a custom model with the Text Generation target type, scroll to Evaluation and moderation and click Configure:

In the Configure evaluation and moderation panel, click a metric card to configure the required settings for the evaluation metric:

If you enabled Use metric as guard, In the Moderation section, configure the moderation settings:

Next, click Add to return to the evaluation selection page. From there, add another metric or click Save configuration.

For a more detailed walkthrough, see [Configure evaluation and moderation](https://docs.datarobot.com/en/docs/workbench/nxt-registry/nxt-model-workshop/nxt-configure-evaluation-moderation.html).

After adding guardrails to a text generation custom model, you can [test](https://docs.datarobot.com/en/docs/workbench/nxt-registry/nxt-model-workshop/nxt-test-custom-model.html), [register](https://docs.datarobot.com/en/docs/workbench/nxt-registry/nxt-model-directory/nxt-register-cus-models.html), and [deploy](https://docs.datarobot.com/en/docs/workbench/nxt-registry/nxt-model-directory/nxt-deploy-models.html) the model to make predictions in production. After you make [predictions](https://docs.datarobot.com/en/docs/workbench/nxt-console/nxt-predictions/index.html), you can view the evaluation metrics on the [Custom metrics](https://docs.datarobot.com/en/docs/workbench/nxt-console/nxt-monitoring/nxt-custom-metrics.html) tab and prompts, responses, and feedback (if configured) on the [Data exploration](https://docs.datarobot.com/en/docs/workbench/nxt-console/nxt-monitoring/nxt-data-exploration.html) tab.

## Define runtime parameters

In the Runtime parameters section, create runtime parameters to provide environment variables to a model at runtime. In addition, if you [defined any runtime parameters](https://docs.datarobot.com/en/docs/api/code-first-tools/drum/custom-model-runtime-parameters.html) through `runtimeParameterDefinitions` in the `model-metadata.yaml` file, you can manage them here. Parameters are injected into containers in two ways:

1. As standard environment variables without prefixes or JSON parsing for simple types (so you can use os.getenv to access environment variables without the datarobot-drum library).
2. For backward compatibility also in the legacy prefixed ( MLOPS_RUNTIME_PARAM_* ) and JSONified format.

Parameters created via the UI persist and merge when you upload new code versions, ensuring a seamless development flow.

**Runtime parameter considerations**

The system uses a blocklist of reserved patterns (e.g., `DRUM_*`, `MLOPS_*`, `KUBERNETES_*`), managed in the dynamic configuration. Matching supports the `*` wildcard (not full regular expression syntax). Reserved names aren't blocked entirely: if a runtime parameter uses a reserved name, the UI displays a warning. How the variable is exposed depends on the context:

- Custom models: Only in prefixed ( MLOPS_RUNTIME_PARAM_* ) and JSONified format—not as a raw (unprefixed) environment variable, to prevent system conflicts.
- Custom apps: Prefixed, but values are not packed into a JSON payload (except for credentials).
- Custom jobs: No prefix and no JSON payload (except for credential types); the variable is available as a raw environment variable.

For credential-type runtime parameters, the system automatically unpacks JSON fields into separate environment variables rather than a single string. For example, a credential named `MAIN_AWS_CREDENTIAL` with the following JSON structure:

```
{"awsAccessKeyId": "<your-key-id>", "awsSecretAccessKey": "<your-access-key>"}
```

is unpacked into the following environment variables, combining the parameter name + JSON key, in uppercase:

```
MAIN_AWS_CREDENTIAL_AWS_ACCESS_KEY_ID="<your-key-id>"
MAIN_AWS_CREDENTIAL_AWS_SECRET_ACCESS_KEY="<your-access-key>"
```

For single-field credential types (for example, `api_token`, `bearer`, or `gcp`), the injected environment variable uses the bare runtime parameter name ( `MY_CRED`), not the parameter name plus the credential field name (for example, not `MY_CRED_API_TOKEN`).Multi-field credential types (for example, `basic` or `s3`) keep the existing suffixed behavior: one variable per field, named `{PARAMETER_NAME}_{FIELD_NAME}` in uppercase snake case (for example, `MY_CRED_USERNAME` and `MY_CRED_PASSWORD`, or the AWS keys in the example above). JSON-encoded runtime parameter variables (for example, `MLOPS_RUNTIME_PARAMETERS_OPEN_AI_API`) are unchanged; only the flat variable for a single-field secret uses the bare parameter name.

> [!TIP] Access runtime parameters in containers
> For programmatic access to runtime parameters in containers, use `DataRobotAppFrameworkBaseSettings` as documented in the SDK API reference.

### Add parameters

To add a new runtime parameter definition, click + Add runtime parameter.

In the Add new runtime parameter dialog box, configure the runtime parameter Name, Type, Value, and, optionally, provide a Description.

After configuring the required fields, click Add, repeat the process for any additional parameters, then click Save.

### Edit parameters

To edit a runtime parameter definition, click Edit.

In the Runtime parameter table, define a runtime parameter Value or, in the last column, access the action icons.

| Icon | Setting | Description |
| --- | --- | --- |
|  | Revert changes | Reset the runtime parameter's Value to the defaultValue set in the model-metadata.yaml file, or to an empty state. |
|  | Delete | Delete the runtime parameter definition. |

If any runtime parameters have `allowEmpty: false` in the definition without a `defaultValue`, you must set a value before registering the custom model.

> [!NOTE] Premium: DataRobot LLM gateway access
> To use the LLM gateway for a custom model with the agentic workflow target type, the `ENABLE_LLM_GATEWAY_INFERENCE` runtime parameter must be provided in the `model-metadata.yaml` file and set to `true`.

For more information on how to define runtime parameters and use them in your custom model code, see the [Define custom model runtime parameters](https://docs.datarobot.com/en/docs/api/code-first-tools/drum/custom-model-runtime-parameters.html) documentation.

## Assign custom model datasets

To enable feature drift tracking for a model deployment, you must add training data. To do this, assign training data to a model version. The method for providing training and holdout datasets for [unstructuredcustom inference models](https://docs.datarobot.com/en/docs/api/code-first-tools/drum/unstructured-custom-models.html) requires you to upload the training and holdout datasets separately. Additionally, these datasets cannot include a partition column.

> [!WARNING] File size warning
> The file size limit for custom model training data uploaded to DataRobot is 1.5GB.

> [!WARNING] Considerations for training data prediction rows count
> Training data uploaded to a custom model is used to compute Feature Impact, drift baselines, and Prediction Explanation previews. To perform these calculations, DataRobot automatically splits the uploaded training data into partitions for training, validation, and holdout (i.e., T/V/H) in a 60/20/20 ratio. Alternatively, you can manually provide a partition column in the training dataset to assign predictions, row-by-row, to the training ( `T`), validation ( `V`), or holdout ( `H`) partitions.
> 
> Prediction Explanations require 100 rows in the validation partition, which—if you don’t define your own partitioning—requires the provided training dataset to contain a minimum of 500 rows. If the training data and partition ratio (defined automatically or manually) result in a validation partition containing fewer than 100 rows, Prediction Explanations are not calculated. While you can still register and deploy the model—and the deployment can make predictions—if you request predictions with explanations, the deployment returns an error.

To assign training data to a custom model after you [define theEnvironmentandFiles](https://docs.datarobot.com/en/docs/workbench/nxt-registry/nxt-model-workshop/nxt-create-custom-model.html#assemble-the-custom-model):

1. On theAssembletab, in theSettingssection, underDatasets:
2. In theTraining Datasection, click+ Select datasetand do either of the following: Include features required for scoringThe columns in a custom model's training data indicate which features are included in scoring requests to the deployed custom model. Therefore, once training data is available, any features not included in the training dataset aren't sent to the model. Available as a preview feature, you can disable this behavior using theColumn filtering setting.
3. (Optional) In thePartition columnsection, specify the column name (from the provided training dataset) containing partitioning information for your data (based on training/validation/holdout partitioning). If you plan to deploy the custom model and monitor itsdata driftandaccuracy, specify the holdout partition in the column to establish an accuracy baseline.
4. ClickAssign. Training data assignment errorIf the training data assignment fails, an error message appears in the new custom model version underDatasets. While this error is active, you can't create a model package to deploy the affected version. To resolve the error and deploy the model package, reassign training data to create a new version, or create a new version andthenassign training data.

### Disable column filtering for prediction requests

> [!NOTE] Preview
> Configurable column filtering is off by default. Contact your DataRobot representative or administrator for information on enabling this feature.
> 
> Feature flag: Enable Feature Filtering for Custom Model Predictions

Now available for preview, you can enable or disable column filtering for custom model predictions. The filtering setting you select is applied in the same way during custom model [testing](https://docs.datarobot.com/en/docs/workbench/nxt-registry/nxt-model-workshop/nxt-test-custom-model.html) and deployment. By default, the target column is filtered out of prediction requests and, if training data is assigned, any additional columns not present in the training dataset are filtered out of any scoring requests sent to the model. Alternatively, if the prediction dataset is missing columns, an error message appears to notify you of the missing features.

You can disable this column filtering when you assemble a custom model. In the Workshop, open a custom model to the Assemble tab, and, in the Settings section, under Column filtering, clear Exclude target column from requests (or, if training data is assigned, clear Exclude target and extra columns not in training data):

As with other changes to a model's environment, files, or settings, changing this setting creates a new [minor custom model version](https://docs.datarobot.com/en/docs/workbench/nxt-registry/nxt-model-workshop/nxt-custom-model-versions.html).

The following behavior is expected when Exclude target column from requests / Exclude target and extra columns not in training data is enabled or disabled:

> [!WARNING] Training data assignment method
> If a model uses [the deprecated "per model" training data assignment method](https://docs.datarobot.com/en/docs/release/deprecations-and-migrations/cus-model-training-data.html), this setting cannot be disabled and feature filtering is not applied during [testing](https://docs.datarobot.com/en/docs/workbench/nxt-registry/nxt-model-workshop/nxt-test-custom-model.html).

| Set to | Behavior |
| --- | --- |
| Enabled | Extra columns: The target column and any extra columns (not present in the training dataset) are excluded from the prediction request.Missing columns: An error message appears to notify you of missing features. |
| Disabled | Extra columns: The columns submitted in the prediction request are not modified.Missing columns: No error message appears to notify you of missing features. |

> [!NOTE] DRUM predictions
> Predictions made with [DRUM](https://docs.datarobot.com/en/docs/api/code-first-tools/drum/custom-model-drum.html) are not filtered; all columns are included in each prediction request.

## Configure custom model resource settings

After creating a custom inference model, you can configure the resources the model consumes to facilitate smooth deployment and minimize potential environment errors in production.

To configure the resource allocation and access settings:

1. On theAssembletab, in theSettingssection, next toResources, click ()Edit:
2. In theUpdate resource settingsdialog box, configure the following settings: Resource settings accessUsers can determine the maximum memory allocated for a model, but onlyorganization adminscan configure additional resource settings. SettingDescriptionMemoryDetermines the maximum amount of memory that can be allocated for a custom inference model. If a model is allocated more than the configured maximum memory value, it is evicted by the system. If this occurs during testing, the test is marked as a failure. If this occurs when the model is deployed, the model is automatically launched again by Kubernetes.ReplicasSets the number of replicas executed in parallel to balance workloads when a custom model is running. Increasing the number of replicas may not result in better performance, depending on the custom model's speed.Network accessPremium feature. Configures the egress traffic of the custom model:Public: The default setting. The custom model can access any fully qualified domain name (FQDN) in a public network to leverage third-party services.None: The custom model is isolated from the public network and cannot access third-party services.When public network access is enabled, your custom model can use theDATAROBOT_ENDPOINTandDATAROBOT_API_TOKENenvironment variables. These environment variables are available for any custom model using adrop-in environmentor acustom environmentbuilt onDRUM. Imbalanced memory settingsDataRobot recommends configuring resource settings only when necessary. When you configure theMemorysetting above, you set the Kubernetes memory "limit" (the maximum allowed memory allocation); however, you can't set the memory "request" (the minimum guaranteed memory allocation). For this reason, it is possible to set the "limit" value too far above the default "request" value. An imbalance between the memory "request" and the memory usage allowed by the increased "limit" can result in the custom model exceeding the memory consumption limit. As a result, you may experience unstable custom model execution due to frequent eviction and relaunching of the custom model. If you require an increasedMemorysetting, you can mitigate this issue by increasing the "request" at the organization level; for more information, contact DataRobot Support. Premium feature: Network accessEverynewcustom model you create haspublic network accessby default; however, when you create new versions of any custom model created before October 2023, those new versions remain isolated from public networks (access set toNone) until you enable public access for a new version (access set toPublic). From this point on, each subsequent version inherits the public access definition from the previous version.
3. Once you have configured the resource settings for the custom model, clickSave. This creates a new minor version of the custom model with edited resource settings applied.

### Select a resource bundle

> [!NOTE] Preview
> Custom model resource bundles and GPU resource bundles are off by default. Contact your DataRobot representative or administrator for information on enabling this feature.
> 
> Feature flag: Enable Resource Bundles, Enable Custom Model GPU Inference ( Premium feature)

You can select a Resource bundle —instead of Memory —when you assemble a model and [configure the resource settings](https://docs.datarobot.com/en/docs/workbench/nxt-registry/nxt-model-workshop/nxt-create-custom-model.html#configure-custom-model-resource-settings). Resource bundles allow you to choose from various CPU and GPU hardware platforms for building and testing custom models. In a custom model's Settings section, open the Resources settings to select a resource bundle. In this example, the model is built to be tested and deployed on an NVIDIA A10 device.

Click Edit to open the Update resource settings dialog box and, in the resource Bundle field, review the CPU and NVIDIA GPU devices available as build environments:

DataRobot can deploy models onto any of the following NVIDIA resource bundles:

| Bundle | GPU | VRAM | CPU | RAM |
| --- | --- | --- | --- | --- |
| GPU - S | 1x NVIDIA T4 | 16GB | 4 | 16GB |
| GPU - M | 1x NVIDIA T4 | 16GB | 8 | 32GB |
| GPU - L | 1x NVIDIA A10G | 24GB | 8 | 32GB |
| GPU - XL | 1x NVIDIA L40S | 48GB | 4 | 32GB |
| GPU - XXL | 4x NVIDIA A10G | 96GB | 48 | 192GB |
| GPU - 3XL | 4x NVIDIA L40S | 192GB | 48 | 384GB |
| GPU - 4XL | 8x NVIDIA A10G | 192GB | 192 | 768GB |
| GPU - 5XL | 8x NVIDIA L40S | 384GB | 192 | 1.5TB |

Along with the NVIDIA GPU resource bundles, this feature introduces custom model environments optimized to run on NVIDIA GPUs. When you assemble a custom model, you define a Base Environment. The model in the example below is running on an [\[NVIDIA\] Triton Inference Server](https://developer.nvidia.com/triton-inference-server):

## Assemble an anomaly detection model

You can create custom models that support anomaly detection problems. If you choose to build one, reference the [DRUM template](https://github.com/datarobot/datarobot-user-models/tree/master/model_templates/python3_sklearn_anomaly). (Log in to GitHub before clicking this link.) When deploying custom anomaly detection models, note that the following functionality is not supported:

- Data drift
- Accuracy and association IDs
- Challenger models
- Humility rules
- Prediction intervals

## Assemble a text generation model

> [!NOTE] Availability information
> Monitoring support for generative models is a premium feature. Contact your DataRobot representative or administrator for information on enabling this feature.

Text generation custom models custom models require the [DataRobot] Python 3.X GenAI drop-in environment, or a compatible [custom environment](https://docs.datarobot.com/en/docs/workbench/nxt-registry/nxt-environment-workshop/index.html). To assemble, test, and deploy a generative model from the workshop, a basic LLM assembled in the model workshop should, at minimum, include the following files:

| File | Contents |
| --- | --- |
| custom.py | The custom model code, implementing the Bolt-on Governance API (the chat() hook) calling the LLM service's API through public network access for custom models. |
| model-metadata.yaml | The custom model metadata and runtime parameters required by the generative model. |
| requirements.txt | The libraries (and versions) required by the generative model. |

After you add the required model files, [add training data](https://docs.datarobot.com/en/docs/workbench/nxt-registry/nxt-model-workshop/nxt-create-custom-model.html#assign-custom-model-datasets). To provide a training baseline for drift monitoring, upload a dataset containing at least 20 rows of prompts and responses relevant to the topic your generative model is intended to answer questions about. These prompts and responses can be taken from documentation, manually created, or generated.

## Assemble a vector database

> [!NOTE] Premium
> Vector database deployments are a premium feature, off by default, and require GenAI experimentation. Contact your DataRobot representative or administrator for information on enabling this feature.

Vector database custom models require the [DataRobot] Python 3.X GenAI drop-in environment, or a compatible [custom environment](https://docs.datarobot.com/en/docs/workbench/nxt-registry/nxt-environment-workshop/index.html). In addition, you must define the dependencies for the vector database. To populate the Dependencies section, you can upload a `requirements.txt` file in the Files section, allowing DataRobot to build the optimal image. In addition, vector database custom models require [public network access](https://docs.datarobot.com/en/docs/workbench/nxt-registry/nxt-model-workshop/nxt-create-custom-model.html#configure-custom-model-resource-settings).

After creating a vector database in the workshop, you can [register](https://docs.datarobot.com/en/docs/workbench/nxt-registry/nxt-model-directory/nxt-register-cus-models.html) and [deploy](https://docs.datarobot.com/en/docs/workbench/nxt-registry/nxt-model-directory/nxt-deploy-models.html) the model, as you would any other custom model.

> [!NOTE] What monitoring is available for vector database deployments?
> DataRobot automatically generates custom metrics relevant to vector databases for deployments with the Vector database deployment type; for example, Total Documents, Average Documents, Total Citation Tokens, Average Citation Tokens, and VDB Score Latency. Vector database deployments also support service health monitoring. Vector database deployments don't store prediction row-level data for data exploration.

## Assemble an agentic workflow

> [!NOTE] Premium
> Agentic workflows are a premium feature. Contact your DataRobot representative or administrator for information on enabling the feature.

Agentic workflows require the [DataRobot] Python 3.X GenAI drop-in environment, or a compatible [custom environment](https://docs.datarobot.com/en/docs/workbench/nxt-registry/nxt-environment-workshop/index.html). To assemble, test, and deploy an agentic workflow from the workshop, a basic agent assembled in the workshop usually includes the following files:

| File | Contents |
| --- | --- |
| custom.py | The custom model code, implementing the Bolt-on Governance API (the chat() hook) to call the LLM and also passing those parameters to the agent (defined in agent.py). |
| agent.py | The agent code, implementing the agentic workflow. |
| model-metadata.yaml | The custom model metadata and runtime parameters required by the agentic workflow. |
| requirements.txt | The libraries (and versions) required by the agentic workflow. |

After creating the custom agentic workflow, you can [test the custom workflow's implementation](https://docs.datarobot.com/en/docs/workbench/nxt-registry/nxt-model-workshop/nxt-test-custom-model.html#__tabbed_1_2) of the [Bolt-on Governance API](https://docs.datarobot.com/en/docs/agentic-ai/genai-code/genai-chat-completion-api.html) (the [chat()hook](https://docs.datarobot.com/en/docs/api/code-first-tools/drum/structured-custom-models.html#chat))

> [!NOTE] Ensure agent tools are accessible
> To deploy a functional agentic workflow, ensure the deployed tools required by the custom agentic workflow are accessible.

> [!NOTE] Agentic workflow deployment functionality
> Agentic workflows can only be deployed on Serverless prediction environments. After deployment, Console includes the deployment Overview tab; Monitoring for service health, usage (including quota monitoring), custom metrics, data exploration (including tracing), resource monitoring, OpenTelemetry (OTel) metrics, and deployment reports; Predictions through the Prediction API (including chat completions); and Activity log for standard output, OTel logs, and moderation events when evaluation and moderation guardrails are configured. For a tab-by-tab summary with links, see [Monitor workflows](https://docs.datarobot.com/en/docs/agentic-ai/agentic-monitor/index.html).

## Register an agentic tool

Agentic tools are registered in the Registry so that they can be easily listed and called by agents, granting them access to the DataRobot API. These tools can be viewed in the Tools tab of the Workshop.

Click the tool to view its details or click Add a tool to register a new tool. This opens the Add a model page, where you can configure the model as you would for any other custom model, with the addition that the Use model as tool box under Optional settings is checked.

Click Add model to register the tool. The tool version opens on the Workshop > Tools tab with its details view.

From here, configure the tool as you would for any other custom model, as described in [Assemble the custom model](https://docs.datarobot.com/en/docs/workbench/nxt-registry/nxt-model-workshop/nxt-create-custom-model.html#assemble-the-custom-model). For details on managing agentic tools, see the [View and manage tools](https://docs.datarobot.com/en/docs/workbench/nxt-registry/nxt-tools/nxt-view-manage-tools.html) page.