> ## Documentation Index
> Fetch the complete documentation index at: https://docs.kiteml.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Create Training Runs

> Launch one run per policy. Returns a list of resources, one per policy.

Everything is validated across the whole batch before anything launches, so a
request either starts all of its runs or none of them.



## OpenAPI

````yaml /api-reference/openapi.json post /v1/training_runs
openapi: 3.1.0
info:
  title: Kite Platform API
  description: >-
    Public API for Kite. Authenticate with `Authorization: Bearer kite_...`, and
    send `Kite-Version: 2026-09-27` to pin the dated contract your code was
    written against (without it you get 2026-09-27). Every response carries the
    version that served it.
  version: '2026-09-27'
servers:
  - url: https://api.kiteml.com
    description: Kite Platform API
security:
  - bearerAuth: []
tags:
  - name: Augmentations
    description: Generate new episodes from a LeRobot dataset.
  - name: Training runs
    description: Fine-tune a policy on a dataset, one run per policy.
  - name: Twins
    description: Rebuild a robot episode as an interactable MuJoCo scene (private beta).
  - name: Catalog
    description: Policies, hardware tiers, and dataset inspection.
  - name: Operations
    description: Poll any run by its id.
  - name: Webhooks
    description: Endpoints that receive events, and the event log.
  - name: API keys
    description: The key making the request.
  - name: Usage
    description: Token usage and credit balance.
paths:
  /v1/training_runs:
    post:
      tags:
        - Training runs
      summary: Create Training Runs
      description: >-
        Launch one run per policy. Returns a list of resources, one per policy.


        Everything is validated across the whole batch before anything launches,
        so a

        request either starts all of its runs or none of them.
      operationId: create_training_runs_v1_training_runs_post
      parameters:
        - name: Idempotency-Key
          in: header
          required: false
          schema:
            anyOf:
              - type: string
                maxLength: 255
              - type: 'null'
            title: Idempotency-Key
        - $ref: '#/components/parameters/KiteVersion'
      requestBody:
        required: true
        content:
          application/json:
            schema:
              $ref: '#/components/schemas/TrainingRunCreate'
      responses:
        '202':
          description: >-
            One run per policy, in the order given: `processing`, or `failed`
            with `submission_failed` when its GPU job could not start
          content:
            application/json:
              schema:
                $ref: '#/components/schemas/TrainingRunList'
        '400':
          description: >-
            Invalid request (`parameter_invalid`, `unsupported_version`, ...);
            `param` names the field
          content:
            application/json:
              schema:
                $ref: '#/components/schemas/ApiErrorResponse'
        '401':
          description: Missing or invalid API key (`unauthenticated`)
          content:
            application/json:
              schema:
                $ref: '#/components/schemas/ApiErrorResponse'
        '402':
          content:
            application/json:
              schema:
                $ref: '#/components/schemas/ApiErrorResponse'
          description: Not enough credits for this run (`insufficient_tokens`)
        '403':
          description: >-
            The key lacks a scope this route needs (`missing_scope`), or the
            feature is not enabled for the account
          content:
            application/json:
              schema:
                $ref: '#/components/schemas/ApiErrorResponse'
        '409':
          content:
            application/json:
              schema:
                $ref: '#/components/schemas/ApiErrorResponse'
          description: >-
            Conflicts with the resource's state or an earlier request (e.g.
            `not_cancelable`, `idempotency_key_reused`)
        '429':
          description: >-
            Rate or concurrency limited (`rate_limit_exceeded`,
            `concurrency_limit_exceeded`, `capacity_exceeded`); honor
            `Retry-After`
          content:
            application/json:
              schema:
                $ref: '#/components/schemas/ApiErrorResponse'
        '503':
          content:
            application/json:
              schema:
                $ref: '#/components/schemas/ApiErrorResponse'
          description: >-
            Temporarily unavailable (`service_unavailable`); retry after
            `Retry-After`
components:
  parameters:
    KiteVersion:
      name: Kite-Version
      in: header
      required: false
      description: >-
        The dated contract your code was written against. Without it you get
        2026-09-27; every response says which version served it.
      schema:
        type: string
        enum:
          - '2026-09-27'
        default: '2026-09-27'
  schemas:
    TrainingRunCreate:
      properties:
        dataset:
          $ref: '#/components/schemas/TrainingRunDataset'
        policies:
          items:
            type: string
          type: array
          maxItems: 8
          minItems: 1
          title: Policies
          description: >-
            LeRobot policy types to train; one run per policy. See GET
            /v1/training_policies. Note: all runs count against your account's
            in-flight concurrency limit, so a batch larger than that limit is
            rejected upfront.
        hardware_tier:
          type: string
          maxLength: 64
          title: Hardware Tier
          description: See GET /v1/hardware_tiers
          default: gcp_gpu_t4
        robot:
          anyOf:
            - type: string
              maxLength: 64
            - type: 'null'
          title: Robot
          description: LeRobot robot tag (label only)
        config:
          $ref: '#/components/schemas/TrainingRunConfig'
        cameras:
          anyOf:
            - $ref: '#/components/schemas/TrainingRunCameras'
            - type: 'null'
        output:
          $ref: '#/components/schemas/TrainingRunOutput'
        webhook_metadata:
          anyOf:
            - type: object
            - type: 'null'
          title: Webhook Metadata
          description: Opaque metadata echoed on the resource and its webhook events
      type: object
      required:
        - dataset
        - policies
      title: TrainingRunCreate
    TrainingRunList:
      properties:
        object:
          type: string
          enum:
            - list
          const: list
          title: Object
        data:
          items:
            $ref: '#/components/schemas/TrainingRun'
          type: array
          title: Data
        has_more:
          type: boolean
          title: Has More
        next_cursor:
          anyOf:
            - type: string
            - type: 'null'
          title: Next Cursor
          description: Pass as `after` for the next page (the last run's `id`)
      type: object
      required:
        - object
        - data
        - has_more
        - next_cursor
      title: TrainingRunList
    ApiErrorResponse:
      properties:
        error:
          $ref: '#/components/schemas/ApiErrorBody'
      type: object
      required:
        - error
      title: ApiErrorResponse
      description: Every /v1 error has this shape.
    TrainingRunDataset:
      properties:
        uri:
          type: string
          maxLength: 500
          minLength: 3
          title: Uri
          description: >-
            LeRobot dataset: a HuggingFace repo id ('lerobot/pusht') or a gs://
            path
      type: object
      required:
        - uri
      title: TrainingRunDataset
    TrainingRunConfig:
      properties:
        steps:
          anyOf:
            - type: integer
              maximum: 1000000
              minimum: 1
            - type: 'null'
          title: Steps
          description: Override the policy's default
        batch_size:
          anyOf:
            - type: integer
              maximum: 1024
              minimum: 1
            - type: 'null'
          title: Batch Size
          description: Override the policy's default
        save_freq:
          anyOf:
            - type: integer
              minimum: 1
            - type: 'null'
          title: Save Freq
          description: Save a checkpoint every N steps
        eval_enabled:
          type: boolean
          title: Eval Enabled
          default: false
      type: object
      title: TrainingRunConfig
    TrainingRunCameras:
      properties:
        rename:
          anyOf:
            - additionalProperties:
                type: string
              type: object
            - type: 'null'
          title: Rename
          description: >-
            Map dataset camera keys to policy image slots; omit to let Kite map
            them
      type: object
      title: TrainingRunCameras
    TrainingRunOutput:
      properties:
        push_to_hub:
          type: boolean
          title: Push To Hub
          description: Push the trained policy to your connected HuggingFace account
          default: false
        hf_repo_id:
          anyOf:
            - type: string
              maxLength: 200
            - type: 'null'
          title: Hf Repo Id
          description: Policy name; the namespace is always your connected HF account
      type: object
      title: TrainingRunOutput
    TrainingRun:
      properties:
        id:
          type: string
          title: Id
          description: '`trn_` + a time-ordered ULID'
        object:
          type: string
          enum:
            - training_run
          const: training_run
          title: Object
        status:
          type: string
          title: Status
          description: '`processing`, `succeeded`, `failed`, `canceled`; an open set'
        phase:
          anyOf:
            - type: string
            - type: 'null'
          title: Phase
          description: >-
            Finer than `status`, e.g. `provisioning`, `preparing`,
            `downloading_dataset`, `training`, `completed`, `failed`,
            `cancelled`. Only GET /v1/training_runs/{id} reads it: null in list,
            create and cancel responses
        status_message:
          anyOf:
            - type: string
            - type: 'null'
          title: Status Message
          description: '`phase` in words, e.g. `Provisioning GPU…`'
        progress:
          type: number
          maximum: 1
          minimum: 0
          title: Progress
        error:
          anyOf:
            - $ref: '#/components/schemas/TrainingRunError'
            - type: 'null'
          description: Why, when `failed`
        policy:
          type: string
          title: Policy
          description: The LeRobot policy type (see GET /v1/training_policies)
        robot:
          anyOf:
            - type: string
            - type: 'null'
          title: Robot
          description: The robot tag from create (a label only)
        hardware_tier:
          type: string
          title: Hardware Tier
          description: See GET /v1/hardware_tiers
        group_id:
          type: string
          title: Group Id
          description: Shared by the runs one create started, one per policy
        dataset:
          $ref: '#/components/schemas/TrainingRunDataset'
        config:
          $ref: '#/components/schemas/TrainingRunResolvedConfig'
        cameras:
          $ref: '#/components/schemas/TrainingRunResolvedCameras'
        metrics:
          $ref: '#/components/schemas/TrainingRunMetrics'
          description: >-
            Live from the trainer on GET /v1/training_runs/{id}; elsewhere, the
            last values stored
        eval:
          anyOf:
            - $ref: '#/components/schemas/TrainingRunEval'
            - type: 'null'
          description: >-
            Once the trainer has reported on evaluation (a finished run has);
            null in list, create and cancel responses
        output:
          $ref: '#/components/schemas/TrainingRunOutputSummary'
          description: What the run has produced
        tokens:
          $ref: '#/components/schemas/TrainingRunTokens'
        webhook_metadata:
          anyOf:
            - type: object
            - type: 'null'
          title: Webhook Metadata
          description: Echoed from create
        created_at:
          type: string
          title: Created At
        started_at:
          anyOf:
            - type: string
            - type: 'null'
          title: Started At
          description: When the GPU job was submitted
        completed_at:
          anyOf:
            - type: string
            - type: 'null'
          title: Completed At
          description: When it became `succeeded`, `failed` or `canceled`
      type: object
      required:
        - id
        - object
        - status
        - phase
        - status_message
        - progress
        - error
        - policy
        - robot
        - hardware_tier
        - group_id
        - dataset
        - config
        - cameras
        - metrics
        - eval
        - output
        - tokens
        - webhook_metadata
        - created_at
        - started_at
        - completed_at
      title: TrainingRun
      description: One policy trained on one GPU.
      examples:
        - cameras: {}
          completed_at: '2026-09-27T15:41:52.664301Z'
          config:
            batch_size: 8
            eval_enabled: false
            save_freq: 1000
            steps: 30000
          created_at: '2026-09-27T14:02:11.482913Z'
          dataset:
            uri: lerobot/svla_so101_pickplace
          eval: {}
          group_id: 5f0c2e9a7b1d4c3e8f6a2b9d0c1e7f35
          hardware_tier: gcp_gpu_t4
          id: trn_01K65Z3N8Q4W2H7R9T6Y1B3C5D
          metrics:
            loss: 0.043
            max_steps: 30000
            step: 30000
          object: training_run
          output:
            checkpoint_count: 30
            hf_push_status: ok
            hf_repo_id: your-hf-user/act-pickplace
            hf_url: https://huggingface.co/your-hf-user/act-pickplace
            latest_checkpoint:
              artifact_uri: >-
                gs://example-bucket/jobs/7c1e9f0a2b3d4e5f8a9b0c1d2e3f4a5b/artifacts/lerobot/checkpoints/030000/pretrained_model/
              created_at: '2026-09-27T15:38:40.120394Z'
              download_url: >-
                /v1/training_runs/trn_01K65Z3N8Q4W2H7R9T6Y1B3C5D/checkpoints/30000/download
              object: training_checkpoint
              step: 30000
          phase: completed
          policy: act
          progress: 1
          robot: so101_follower
          started_at: '2026-09-27T14:02:13.107554Z'
          status: succeeded
          status_message: Completed
          tokens:
            charged: 500
            refunded: 0
          webhook_metadata:
            ref: nightly-42
    ApiErrorBody:
      properties:
        type:
          type: string
          title: Type
          description: >-
            invalid_request_error, authentication_error, permission_error,
            not_found_error, conflict_error, rate_limit_error, or api_error
        code:
          type: string
          title: Code
          description: 'Machine-readable: branch on this, not on the message'
        message:
          type: string
          title: Message
        param:
          anyOf:
            - type: string
            - type: 'null'
          title: Param
          description: The request field the error is about, when there is one
        request_id:
          anyOf:
            - type: string
            - type: 'null'
          title: Request Id
          description: Also in the X-Request-Id header; quote it to support
        details:
          anyOf:
            - type: object
            - type: 'null'
          title: Details
          description: >-
            Extra data for some codes (e.g. the cameras a dataset does have, for
            camera_not_found)
      type: object
      required:
        - type
        - code
        - message
      title: ApiErrorBody
    TrainingRunError:
      properties:
        code:
          type: string
          title: Code
          description: >-
            `training_failed`, `submission_failed`, `vertex_state_unobserved`,
            `service_unavailable`, `gpu_capacity_unavailable`.
            `submission_failed`: the GPU job never started, so nothing was
            charged
        message:
          type: string
          title: Message
          description: What went wrong, in at most 1,000 characters
      type: object
      required:
        - code
        - message
      title: TrainingRunError
    TrainingRunResolvedConfig:
      properties:
        steps:
          type: integer
          title: Steps
          description: '`config.steps` from create, else the policy''s default'
        batch_size:
          type: integer
          title: Batch Size
          description: '`config.batch_size` from create, else the policy''s default'
        save_freq:
          type: integer
          title: Save Freq
          description: A checkpoint is saved every this many steps
        eval_enabled:
          type: boolean
          title: Eval Enabled
      type: object
      required:
        - steps
        - batch_size
        - save_freq
        - eval_enabled
      title: TrainingRunResolvedConfig
    TrainingRunResolvedCameras:
      properties:
        rename:
          anyOf:
            - additionalProperties:
                type: string
              type: object
            - type: 'null'
          title: Rename
          description: >-
            Dataset camera key → policy image slot, from create; null when Kite
            maps them
      type: object
      required:
        - rename
      title: TrainingRunResolvedCameras
    TrainingRunMetrics:
      properties:
        step:
          anyOf:
            - type: integer
            - type: 'null'
          title: Step
          description: The last step the trainer reported
        max_steps:
          anyOf:
            - type: integer
            - type: 'null'
          title: Max Steps
        loss:
          anyOf:
            - type: number
            - type: 'null'
          title: Loss
          description: The training loss at `step`
      type: object
      required:
        - step
        - max_steps
        - loss
      title: TrainingRunMetrics
    TrainingRunEval:
      properties:
        status:
          anyOf:
            - type: string
            - type: 'null'
          title: Status
          description: >-
            `enabled`, `ok` or `skipped: <reason>`; null when
            `config.eval_enabled` is false
        metrics:
          anyOf:
            - additionalProperties:
                type: number
              type: object
            - type: 'null'
          title: Metrics
          description: Rollout results, e.g. `pc_success`, `avg_sum_reward`
      type: object
      required:
        - status
        - metrics
      title: TrainingRunEval
    TrainingRunOutputSummary:
      properties:
        checkpoint_count:
          type: integer
          title: Checkpoint Count
          description: >-
            Checkpoints saved so far. Only GET /v1/training_runs/{id} reads
            storage: it is 0 in list, create and cancel responses
        latest_checkpoint:
          anyOf:
            - $ref: '#/components/schemas/TrainingCheckpoint'
            - type: 'null'
          description: The newest; all of them are GET /v1/training_runs/{id}/checkpoints
        hf_repo_id:
          anyOf:
            - type: string
            - type: 'null'
          title: Hf Repo Id
          description: >-
            Only when the run pushes to the Hub (`output.push_to_hub` at
            create): the repo it pushes to
        hf_url:
          anyOf:
            - type: string
            - type: 'null'
          title: Hf Url
          description: 'Only when pushing: the repo''s page'
        hf_push_status:
          anyOf:
            - type: string
            - type: 'null'
          title: Hf Push Status
          description: >-
            Only when pushing: null until the push starts, then `pushing
            (attempt n/5)`, `ok`, `skipped: <reason>` or `failed: <reason>`,
            e.g. `skipped: no Hugging Face account connected`
      type: object
      required:
        - checkpoint_count
        - latest_checkpoint
      title: TrainingRunOutputSummary
    TrainingRunTokens:
      properties:
        charged:
          type: integer
          title: Charged
          description: >-
            Reserved at launch: one GPU-hour at the tier's rate (0 on a CPU
            tier, or when the GPU job never started)
        refunded:
          type: integer
          title: Refunded
          description: >-
            Given back when the run finished. 0 today: a run is charged its
            GPU-hour however long it trains
      type: object
      required:
        - charged
        - refunded
      title: TrainingRunTokens
    TrainingCheckpoint:
      properties:
        object:
          type: string
          enum:
            - training_checkpoint
          const: training_checkpoint
          title: Object
        step:
          type: integer
          title: Step
        artifact_uri:
          anyOf:
            - type: string
            - type: 'null'
          title: Artifact Uri
          description: >-
            Where it is stored (`gs://…/pretrained_model/`), for reading the
            bucket directly
        download_url:
          type: string
          title: Download Url
          description: >-
            This API's path to it as a ZIP of `pretrained_model/` (send your
            Authorization header)
        created_at:
          anyOf:
            - type: string
            - type: 'null'
          title: Created At
          description: When it was saved; null for older runs
      type: object
      required:
        - object
        - step
        - artifact_uri
        - download_url
        - created_at
      title: TrainingCheckpoint
      description: A checkpoint a training run saved.
      examples:
        - artifact_uri: >-
            gs://example-bucket/jobs/7c1e9f0a2b3d4e5f8a9b0c1d2e3f4a5b/artifacts/lerobot/checkpoints/030000/pretrained_model/
          created_at: '2026-09-27T15:38:40.120394Z'
          download_url: >-
            /v1/training_runs/trn_01K65Z3N8Q4W2H7R9T6Y1B3C5D/checkpoints/30000/download
          object: training_checkpoint
          step: 30000
  securitySchemes:
    bearerAuth:
      type: http
      scheme: bearer
      description: >-
        A Kite API key (`kite_...`), from **Platform API → Keys** in the
        dashboard.

````