Model properties
The HubContentDocument for the Model HubContentType contains
the metadata that describes a JumpStart model, including its id, licensing, deployment
configuration, and customization options. Use this metadata to evaluate and compare models
and to make informed decisions about which model best fits your use case.
When calling the DescribeHubContent
operation for a model, the response includes a HubContentDocument field,
returned as a JSON-formatted string. This section describes the properties of that document
for the Model HubContentType so that it can be more easily interpreted.
Parse the HubContentDocument string as JSON to read the properties described
here. Property names use PascalCase. A document contains only the properties that apply to
the model, so most properties are optional.
The following properties are always present: Url,
TrainingSupported, IncrementalTrainingSupported,
ModelTypes, and Capabilities. All other properties are optional
and appear only when they apply to the model.
Top-level properties
The following table lists the top-level properties of the model
HubContentDocument.
Property |
Type |
Description |
|---|---|---|
|
|
The URL of the model's source page. Required |
|
|
The oldest Python SDK version compatible with this model. Optional |
|
|
Specifies whether the model supports the classic Python SDK training path (not the recipe-based path). Required |
|
|
Specifies whether the model can be fine-tuned again starting from a prior training job's output artifact. Required |
|
|
The base name used to automatically name endpoints and training jobs. Optional |
|
|
Specifies whether the model weights are gated. To deploy a gated model, you must accept the EULA (accept_eula=True). Optional |
|
|
A marker that identifies Amazon Nova model variants. Optional |
|
|
The licensing class for the model. Valid values are PROPRIETARY (locked weights) or OPEN_WEIGHTS (portable weights). Required |
|
|
The ML task category (for example, text-generation or image-classification). Optional |
|
|
The origin/framework family of the model (for example, huggingface, meta, or nova). Optional |
|
|
The organization that produced the model (for example, Meta, Anthropic, or Amazon). Optional |
|
|
The modality of the model (text, vision, audio, or multimodal). Optional |
|
|
A short license identifier (for example, Apache-2.0 or MIT). Optional |
|
|
The input types that the model accepts (text, image, audio, video, and so on). Optional |
|
|
The output types that the model produces (text, image, audio, video, and so on). Optional |
|
|
An access-tier tag (public, gated, or private-hub). Optional |
|
|
The human languages the model was trained on. Optional |
|
|
The model size by number of parameters. Optional |
|
|
The maximum number of input tokens the model can process. Optional |
|
|
The ML framework the model was originally trained in (for example, PyTorch). Optional |
|
|
The source Hugging Face Hub ID, if the model was mirrored from Hugging Face. Optional |
|
|
A URL to the provider's logo image for the model card. Optional |
|
|
A list of marketing highlights shown on the model detail page. Optional |
|
|
A list of external links (paper, blog, or code repository) shown under "Learn more." Optional |
|
|
The inline help text shown next to the training-data input ( Optional |
|
|
The instructions for copying artifacts into a hub at import time. Optional |
|
|
The container for the example notebooks shown on the model's Notebooks tab. Optional |
|
|
A collection of recipe-based customizations, independent of Optional |
|
|
The feature gates that the model supports (for example, BEDROCK_CONSOLE, TRAINING, FINE_TUNING, CUSTOMIZATION, HYPERPOD_DEPLOYMENT). Required |
|
|
A named set of alternate deploy configurations that override root hosting fields. The top-ranked configuration is the default. Optional |
|
|
A reusable component pool referenced by InferenceConfigs[*].ComponentNames. Optional |
|
|
Specifies which InferenceConfigs entry is the default (overall.Rankings[0]). Optional |
|
|
A named set of training configurations for the classic training path. Optional |
|
|
A reusable component pool referenced by TrainingConfigs[*].ComponentNames. Optional |
|
|
Specifies which TrainingConfigs entry is the default. Optional |
|
|
Specifies whether the model can be deployed as an Inference Component (dynamic-container endpoint). Optional |
|
|
Additional deploy-time model-data channels, typically used for speculative-decoding draft models. Optional |
|
|
The I/O contract ID used when the model is invoked from Amazon Bedrock. Optional |
|
|
The inference container image used for deployment. Optional (effectively required to deploy) |
|
|
The framework/version/Python triple used to compute a Deep Learning Container image. Optional |
|
|
Specifies whether HostingArtifactUri is an S3Prefix (directory) or S3Object (single tarball). Optional |
|
|
Specifies whether HostingArtifactUri is Gzip-compressed or None. Optional |
|
|
The S3 location of the model weights. Optional |
|
|
The S3 location of the inference script package (sourcedir.tar.gz). Optional |
|
|
Specifies whether to fetch a separate script package (false = prepacked). Optional |
|
|
The S3 URI of the model's EULA; its presence gates the model (requires accept_eula). Optional |
|
|
An external link to the vendor's license page. Optional |
|
|
An AWS Marketplace subscribe link. Optional |
|
|
The AWS Marketplace listing ID. Optional |
|
|
The AWS Marketplace product ID. Optional |
|
|
The AWS Marketplace deploy path. The endpoint is created from a model package instead of a container image and artifact. Optional |
|
|
The SageMaker-managed AMI (driver/CUDA version) for the endpoint host. Optional |
|
|
The maximum seconds SageMaker waits for the model weights to download from S3. Optional |
|
|
The maximum seconds SageMaker waits for the container's health check to return 200. Optional |
|
|
The container environment variables set at deploy time. Optional |
|
|
The Python packages that the inference container installs at startup. Optional |
|
|
The default deploy instance type. Optional |
|
|
The instance types available for deployment. Optional |
|
|
The Content-Type and Accept headers used when invoking the model. Optional |
|
|
The additional EBS volume, in GB, attached to the endpoint host. Optional |
|
|
Specifies whether outbound network is blocked from the endpoint container. Optional |
|
|
The canonical example request bodies for the sample-payload pane. Optional |
|
|
The CPU, GPU, and memory requirements for deploying as an Inference Component. Optional |
|
|
The per-instance-family overrides for the deploy configuration (image URIs, environment variables, and model-package ARNs). Optional |
|
|
Specifies whether TrainingArtifactUri is an S3Prefix or S3Object. Optional |
|
|
Specifies whether TrainingArtifactUri is Gzip-compressed or None. Optional |
|
|
The AWS Marketplace training path. Weights come from a model package instead of the model channel. Optional |
|
|
The customer-tunable parameters (name, type, default, minimum, and maximum) for the training path. Optional |
|
|
The S3 location of the training script package (sourcedir.tar.gz). Optional |
|
|
The training container image. Optional |
|
|
The regular expression patterns scraped from logs into Amazon CloudWatch metrics. Optional |
|
|
The pre-trained weights fed into the training job as its starting point. Optional |
|
|
The Python packages that the training container installs at startup. Optional |
|
|
The default training instance type. Optional |
|
|
The instance types available for training. Optional |
|
|
The EBS volume (GB) attached to the training host. Optional |
|
|
Specifies whether outbound network is blocked from the training container. Optional |
|
|
A legacy property. Optional |
|
|
A legacy property. The VALIDATION capability supersedes this property. Optional |
|
|
A sample dataset the wizard can pre-fill for a one-step fine-tune. Optional |
|
|
Specifies whether to encrypt traffic between hosts during distributed training. Optional |
|
|
The maximum number of seconds before the training job is stopped, even if it hasn't converged. Optional |
|
|
Specifies whether to skip compressing the training output artifact. Optional |
|
|
The directory inside the training container where the script writes the final model. Optional |
|
|
The per-instance-family overrides for the training configuration (image URIs and gated-model environment variables). Optional |
Nested properties
The following table lists the properties nested inside the object and array properties in the preceding table. A nested property marked as required is required only when its parent property is present.
Note
The TrainingConfigs and TrainingConfigRankings
properties have the same nested structure as their InferenceConfigs
and InferenceConfigRankings counterparts. Replace
Inference with Training in the property paths.
Property path |
Type |
Description |
|---|---|---|
|
|
The source S3 URI to copy from at ImportHubContent time. Optional |
|
|
The destination S3 URI inside the hub. Optional |
|
|
The kind of artifact being copied. Optional |
|
|
The human-readable link label. Optional |
|
|
The external URL (paper, GitHub, blog). Optional |
|
|
The paragraphs that explain the expected training-data format (up to 50). Optional |
|
|
The paragraphs that describe the sample dataset (up to 50). Optional |
|
|
The legacy single-notebook location. Optional |
|
|
The list of example notebooks shown on the Notebooks tab. Optional |
|
|
The Markdown content for the "Fine-tune" tab overview. Optional |
|
|
The Markdown content for the "Deploy" tab overview. Optional |
|
|
The display title of the notebook. Required |
|
|
The S3 location of the .ipynb file. Required |
|
|
Specifies whether this notebook is pre-selected when the tab opens. Optional |
|
|
The hyperparameter identifier passed to the training script. Required |
|
|
The display label shown in the form. Optional |
|
|
A description or tooltip. Optional |
|
|
The data type of the hyperparameter value. Required |
|
|
The default value; type must match Type. Required |
|
|
The minimum allowed value. Optional |
|
|
The maximum allowed value. Optional |
|
|
The allowed values (text type only). Optional |
|
|
Specifies whether the value is a training-script argument or a container environment variable. Optional |
|
|
The additional format validation applied to the value. Optional |
|
|
The environment variable name. Required |
|
|
The value type. Required |
|
|
The default value; type must match Type. Required |
|
|
A description or tooltip. Optional |
|
|
The allowed values (text type only). Optional |
|
|
The minimum value or length. Optional |
|
|
The maximum value or length. Optional |
|
|
The value scope tag. Optional |
|
|
The model class for which the value is required. Optional |
|
|
The additional format validation applied to the value. Optional |
|
|
The metric name published to CloudWatch. Required |
|
|
The regex applied to logs; captures exactly one numeric group. Required |
|
|
The default request Content-Type header. Required |
|
|
The Content-Types that the model accepts. Required |
|
|
The default response Accept header. Required |
|
|
The Accept types that the model returns. Required |
|
|
The Content-Type header for this example. Required |
|
|
The JSON path where a user prompt is inserted into Body. Optional |
|
|
The example request body. Required |
|
|
The JSON path in the response holding generated text. Optional |
|
|
The JSON path in the response holding input log-probs. Optional |
|
|
The required accelerator count for an Inference Component. Optional |
|
|
The required vCPU count. Optional |
|
|
The required minimum RAM in MB. Required |
|
|
The SageMaker channel name (mount point). Required |
|
|
The version tag for the extra artifact. Optional |
|
|
The Amazon S3 data type (S3Prefix or S3Object). Required |
|
|
The S3 location of the extra artifact. Required |
|
|
The compression type (None or Gzip). Required |
|
|
The HubContent ARN to attribute the download to. Optional |
|
|
The EULA that a user must accept before accessing this additional data source. Optional |
|
|
The provider name for attribution. Required |
|
|
The provider classification. Required |
|
|
An image URI shared across many instance types. Optional |
|
|
A per-instance-type inference image override. Optional |
|
|
The per-instance-type environment variable overrides. Optional |
|
|
A per-instance-type AWS Marketplace model package override. Optional |
|
|
A per-instance-type AWS Marketplace listing ID. Optional |
|
|
A per-instance-type AWS Marketplace product ID. Optional |
|
|
An image URI shared across many instance types. Optional |
|
|
A per-instance-type training image override. Optional |
|
|
The gated-model weights URI passed through an environment variable. Optional |
|
|
A per-instance-type training artifact override. Optional |
|
|
The per-instance-type environment variable overrides. Optional |
|
|
The machine-readable recipe ID, unique within the collection. Required |
|
|
The human-readable recipe title. Required |
|
|
Specifies whether the recipe fine-tunes or evaluates the model. Required |
|
|
The fine-tuning algorithm. Optional |
|
|
The evaluation methodology (only when Type=Evaluation). Optional |
|
|
Specifies whether the recipe requires an active subscription before the model can be customized. Optional |
|
|
The parameter-efficient fine-tuning mode. Optional |
|
|
The training context length for the fine-tune. Optional |
|
|
The accelerator hardware family (for example, GPU, TRAINIUM, or CPU). Optional |
|
|
The model versions the recipe is compatible with. Required |
|
|
The training container image used when the recipe runs as a SageMaker training job. Optional |
|
|
The default number of instances for the training job. Optional |
|
|
The training framework family that runs the recipe. Optional |
|
|
The serverless billing model (per-token or per-hour). Optional |
|
|
The SKU the serverless meter charges against. Optional |
|
|
The deployment targets available after customization. Optional |
|
|
The free-text use-case tags. Optional |
|
|
The free-text "best for" tags. Optional |
|
|
A free-text description of the recipe. Optional |
|
|
The training instance types the recipe supports. Optional |
|
|
The path to the recipe definition file. Optional |
|
|
The S3 location of the HyperPod (EKS) launch template. Optional |
|
|
The S3 location of the HyperPod (EKS) parameter overrides. Optional |
|
|
The S3 location of the SageMaker training job launch template. Optional |
|
|
The S3 location of the SageMaker training job parameter overrides. Optional |
|
|
An instance type the customized model can be deployed on. Required |
|
|
An optional deploy profile name. Optional |
|
|
An optional override for the inference container image. Optional |
|
|
The container environment variables for this deploy target. Optional |
|
|
The minimum RAM in MB. Required (if parent object set) |
|
|
The minimum vCPU count. Required (if parent object set) |
|
|
The minimum accelerator count. Required (if parent object set) |
|
|
The components (from InferenceConfigComponents) merged into this config. Required |
|
|
The other HubContent that this configuration depends on. Optional |
|
|
The metric name (throughput, latency, and so on). Required |
|
|
The measured value. Required |
|
|
The unit (tokens/s, ms, and so on). Required |
|
|
A display label for the metric. Optional |
|
|
The concurrency level at which measured. Optional |
|
|
The acceleration type (compiler, kernel fusion, and so on). Required |
|
|
Specifies whether this acceleration is applied. Required |
|
|
The compiler used. Optional |
|
|
The compiler version. Optional |
|
|
The do-it-yourself (DIY) workflow override flag. Optional |
|
|
The reason for the override. Optional |
|
|
A description of the ranking dimension (for example, "cheapest first"). Optional |
|
|
The ordered list of config names for this ranking. Optional |
