--- title: "Creates a training job" method: POST path: "/v1/training_projects/{training_project_id}/jobs" --- # Creates a training job `POST /v1/training_projects/{training_project_id}/jobs` Creates a training job with the specified configuration. ## Request body - CreateTrainingJobRequestV1 — A request to create a training job. - `training_job` CreateTrainingJobV1, required — Configuration for a training job. - `image` CreateTrainingJobImageV1, required — Configuration to create a training job image. - `base_image` string, required — Base image for the training job. - `docker_auth` DockerAuthV1 — Docker authentication credentials. - `registry` string, required — Registry to authenticate with - `auth_method` 'GCP_SERVICE_ACCOUNT_JSON' | 'AWS_IAM' | 'AWS_OIDC' | 'GCP_OIDC' | 'REGISTRY_SECRET', required - `gcp_service_account_json_docker_auth` GcpServiceAccountJsonDockerAuthV1 — GCP details for the registry. - `service_account_json_secret_ref` SecretReferenceV1, required - `name` string, required — Name of the secret to reference. - `aws_iam_docker_auth` AwsIamDockerAuthV1 — AWS details for the registry. - `access_key_secret_ref` SecretReferenceV1, required - `name` string, required — Name of the secret to reference. - `secret_access_key_secret_ref` SecretReferenceV1, required - `name` string, required — Name of the secret to reference. - `aws_oidc_docker_auth` AwsOidcDockerAuthV1 — AWS OIDC details for the registry. - `role_arn` string, required — AWS IAM role ARN for OIDC authentication - `region` string, required — AWS region for OIDC authentication - `gcp_oidc_docker_auth` GcpOidcDockerAuthV1 — GCP OIDC details for the registry. - `service_account` string, required — GCP service account name for OIDC authentication - `workload_identity_provider` string, required — GCP workload identity provider for OIDC authentication - `registry_secret_docker_auth` RegistrySecretDockerAuthV1 — Authentication via a Baseten secret for any Docker registry (Docker Hub, GHCR, NGC, etc.). The referenced secret must contain credentials in the format 'username:password'. For Docker Hub, set registry to 'https://index.docker.io/v1/'. For GHCR, use 'ghcr.io'. - `secret_ref` SecretReferenceV1, required - `name` string, required — Name of the secret to reference. - `compute` CreateTrainingJobComputeV1 — Configuration to specify the compute for a training job. - `node_count` integer — Number of nodes for the training job. - `cpu_count` integer — Number of cpus for the training job. - `memory` string — Memory for the training job. - `accelerator` CreateTrainingJobAcceleratorV1 - `accelerator` string, required — GPU type for the training job. - `count` integer, required — GPUs needed for the training job. - `availability_model` 'dedicated' | 'spot' — Capacity guarantee under which a training job is scheduled. ``DEDICATED`` is on-demand capacity that is not preempted (the default). ``SPOT`` is interruptible capacity that may be preempted; the user is responsible for checkpointing their own progress. A managed/resumable model where the platform handles checkpoint/resume on its own is intentionally not defined yet; it is planned for a future milestone. - `runtime` CreateTrainingJobRuntimeV1 — Configuration to specify the runtime environment for a training job. - `start_commands` string[] — Commands to execute when starting the runtime. - `environment_variables` object — Environment variables to set in the runtime. - `artifacts` CreateTrainingJobS3Artifact[] — Runtime artifacts for the training job. - `s3_bucket` string, required — S3 bucket for the uploaded runtime artifact. - `s3_key` string, required — S3 key for the uploaded runtime artifact. - `enable_cache` boolean, nullable — Deprecated. Use cache_config instead. - `cache_config` CreateTrainingJobCacheConfig - `enable_legacy_hf_mount` boolean — Whether to enable the legacy Hugging Face cache. - `enabled` boolean — Whether to enable the read-write cache. - `require_cache_affinity` boolean — Whether to require region affinity for the read-write cache. If False, the resulting job is not guaranteed to be deployed alongside the previous cache. - `mount_base_path` string — Mount base path for the cache directory. The project cache and team cache will be mounted under this path. - `checkpointing_config` CreateTrainingJobCheckpointingConfig - `enabled` boolean — Whether checkpointing is enabled. - `checkpoint_path` string, nullable — path where checkpoints will be saved. - `volume_size_gib` integer, nullable — Size of the volume in gibibytes. If not provided, the default size will be used - `load_checkpoint_config` LoadCheckpointConfig - `enabled` boolean — Whether checkpoint loading is enabled - `download_folder` string — Folder where checkpoints will be downloaded - `checkpoints` union[] — List of checkpoint configurations - union - BasetenLatestCheckpointConfig - … - BasetenNamedCheckpointConfig - … - LoopsCheckpointConfig - … - `name` string, nullable — Name of the training job. - `truss_user_env` TrussUserEnv — This data models is used to flexibly store info alongside oracle versions. There is a corresponding data model in the truss client. In contrast, here all fields are optional for backwards compatibility with old clients. - `truss_client_version` string, nullable - `python_version` string, nullable - `pydantic_version` string, nullable - `mypy_version` string, nullable - `is_library_deployment` boolean - `is_frontend_deployment` boolean - `git_info` GitInfo - `latest_commit_sha` string, required - `latest_tag` string, nullable, required - `commits_since_tag` integer, nullable, required - `has_uncommitted_changes` boolean, required - `interactive_session` InteractiveSessionConfigV1 — Configuration for interactive debugging sessions on training jobs. - `trigger` 'on_startup' | 'on_failure' | 'on_demand' - `timeout_minutes` integer — Number of minutes before the interactive session times out. - `session_provider` 'vs_code' | 'cursor' | 'ssh' - `auth_provider` 'github' | 'microsoft' - `weights` CreateJobWeightConfigV1[] — MDN weight sources to mount in the training container. Weights are mirrored and cached for fast startup. - `source` string, required — Weight source URI. Supported formats: hf://, s3://, gs://, r2://, cw:// - `mount_location` string, required — Path where weights will be mounted in the container - `allow_patterns` string[], nullable — File patterns to include (Unix-style shell patterns) - `ignore_patterns` string[], nullable — File patterns to exclude (Unix-style shell patterns) - `auth_secret_name` string, nullable — Name of the workspace secret for authentication (e.g., HuggingFace token) - `auth` object, nullable — Authentication configuration for the weight source. - `enable_baseten_workdir` boolean — When enabled, uses /b10/workspace as the working directory instead of the image WORKDIR. - `priority` integer, nullable — Queue priority. Higher values are dequeued first. Defaults to 0. ## Response `200` - CreateTrainingJobResponseV1 — A response to creating a training job. - `training_job` TrainingJobV1, required - `id` string, required — Unique identifier of the training job. - `created_at` string, date-time, required — Time the job was created in ISO 8601 format. - `current_status` string, required — Current status of the training job. - `error_message` string, nullable — Error message if the training job failed. - `instance_type` InstanceTypeV1, required — An instance type. - `id` string, required — Identifier string for the instance type - `name` string, required — Display name of the instance type - `memory_limit_mib` integer, required — Memory limit of the instance type in Mebibytes - `millicpu_limit` integer, required — CPU limit of the instance type in millicpu - `gpu_count` integer, required — Number of GPUs on the instance type - `gpu_type` string, nullable, required — Type of GPU on the instance type - `gpu_memory_limit_mib` integer, nullable, required — Memory limit of the GPU on the instance type in Mebibytes - `updated_at` string, date-time, required — Time the job was updated in ISO 8601 format. - `training_project_id` string, required — ID of the training project. - `training_project` TrainingProjectSummaryV1, required — A summary of a training project. - `id` string, required — Unique identifier of the training project. - `name` string, required — Name of the training project. - `name` string, nullable — Name of the training job. - `checkpoint_sync_status` 'SYNCING' | 'COMPLETED' — Lifecycle state for the checkpoint uploader. - `priority` integer — Queue priority. Higher values are dequeued first. NULL is treated as 0. - `availability_model` 'dedicated' | 'spot' — Capacity guarantee under which a training job is scheduled. ``DEDICATED`` is on-demand capacity that is not preempted (the default). ``SPOT`` is interruptible capacity that may be preempted; the user is responsible for checkpointing their own progress. A managed/resumable model where the platform handles checkpoint/resume on its own is intentionally not defined yet; it is planned for a future milestone. - `user` UserV1 — A user. - `email` string, nullable — Email of the user. --- [API](https://skmtc.net/baseten/apis/baseten-inference-api.md) · [All operations](https://skmtc.net/baseten/apis/baseten-inference-api/llms.txt) · [OpenAPI document](https://skmtc-service-staging.skmtc.workers.dev/v1/apis/baseten/baseten-inference-api/revisions/a55a2f64c3c5/schema)