llama-stack-mirror/synthetic_data_generation.yaml

openapi: 3.0.0
info:
  title: API for Synthetic Data Generation. This combines other serivces like batch inference and reward model scoring.
  version: 0.0.1
paths:
  /synthetic_data_generation/submit_job:
    post:
      summary: Submit a job for synthetic data generation.
      description: Batch Inference > Reward Scoring > Filtering > Response
      requestBody:
        required: true
        content:
          application/json:
            schema:
              type: object
              properties:
                # batch inference params
                model:
                  type: string
                  description: Model identifier for batch inference.
                prompts_path:
                  type: string
                  description: Path to prompts, JSONL for batch inference
                batch_size:
                  type: integer
                  description: Number of prompts to process in each batch.
                # TODO: May-be put all these generation related params in a struct
                temperature:
                  type: number
                  format: float
                  description: Temperature parameter for generation.
                top_p:
                  type: number
                  format: float
                  description: Top-p parameter for generation.
                max_gen_len:
                  type: integer
                  description: Maximum length of generated responses.
                num_generations:
                  type: integer
                  description: Number of generations per prompt.
                # reward model scoring params
                reward_model:
                  type: string
                  description: Identifier for the reward model used for scoring.
                scoring_function:
                  type: string
                  description: Scoring function to apply.
                # params for filtering responses
                # filtering function will have a signature as
                # def filter_responses(List[PromptResponseScore]) --> List[PromptResponseScore]: ...
                filtering_function:
                  type: object
                  properties:
                    name:
                      type: string
                      description: Name of the filtering function, can be a simple threshold or a pre-registered function.
                    params:
                      type: object
                      additionalProperties: true
                      description: JSON object containing parameters for the filtering function.
      responses:
        '200':
          description: Job successfully created and processing.
          content:
            application/json:
              schema:
                $ref: '#/components/schemas/SyntheticDataGeneration'

/synthetic_data_generation/job_status:
    get:
      summary: Get the status of a submitted job
      description: Get the status of a submitted job
      parameters:
        - in: query
          name: job_id
          required: true
          schema:
            type: string
          description: Unique identifier for the job
      responses:
        '200':
          description: Job status retrieved successfully
          content:
            application/json:
              schema:
                $ref: '#/components/schemas/SyntheticDataGeneration'
        '400':
          description: Invalid job ID provided
        '404':
          description: Job not found

components:
  schemas:
    PromptResponseScore:
      type: object
      properties:
        id:
          type: string
          description: Carry forwarded from the user provided id from prompt.
        index:
          type: integer
          description: Index of the generation.
        prompt:
          type: array
          items:
            $ref: '#/components/schemas/Message'
        response:
          $ref: '#/components/schemas/Completion'
        score:
          type: number
          format: float
          description: Final score after filtering.
        raw_score:
          type: number
          format: float
          description: Raw score from the reward model.
    SyntheticDataGeneration:
      type: object
      properties:
        job_id:
          type: string
          description: Unique identifier for the job.
        created:
          type: string
          format: date-time
          description: Timestamp when the job was created.
        status:
          type: string
          description: Current status of the job, can indicate the stage or success/failure.
        output_file_path:
          type: string
          description: Path to the output jsonl file where each row is a json encoded PromptResponseScore object.
    Message:
      type: object
      properties:
        # As Defined in /batch_inference
    Completion:
      type: object
      properties:
        # As Defined in /batch_inference