assets/large_language_models/rag/components/qa_data_generation/spec.yaml

$schema: https://azuremlschemas.azureedge.net/latest/commandComponent.schema.json
type: command

tags:
    Preview: ""

version: 0.0.78
name: llm_rag_qa_data_generation
display_name: LLM - Generate QnA Test Data
is_deterministic: true

description: |
  Generates a test dataset of questions and answers based on the input documents.

  A chunk of text is read from each input document and sent to the specified LLM with a prompt to create a question and answer based on that text. These question, answer, and context sets are saved as either a csv or jsonl file.
  Short-answer, long-answer, summary, and boolean-based QAs are generated.

inputs:
  openai_api_version:
    type: string
    default: "2023-03-15-preview"
    description: "Version of OpenAI API to use for communicating with LLM."
  openai_api_type:
    type: string
    default: "azure"
    description: "Type of OpenAI endpoint hosting model. Defaults to azure for AOAI endpoints."
  input_data:
    type: uri_folder
    description: "Uri folder of documents containing chunks of data."
  llm_config:
    type: string
    default: '{"type": "azure_open_ai", "model_name": "gpt-35-turbo", "deployment_name": "gpt-35-turbo", "temperature": 0, "max_tokens": 2000}'
    description: "JSON Configuration for what model to use for question generation. Must contain following keys: 'type' (value must be 'azure_open_ai' or 'azure'), 'model_name' (name of model to use for summary), 'deployment_name' (name of deployment for model), 'temperature' (randomness in response, float from 0 to 1), 'max_tokens' (number of tokens for response)."
  llm_connection:
    type: string
    optional: False
    description: "Workspace connection resource ID for the completion model."
  dataset_size:
    type: integer
    default: 100
    description: "Number of questions to generate"
  chunk_batch_size:
    type: integer
    default: 5
    description: "Number of chunks to be read and sent to LLM in parallel"
  output_format:
    type: string
    default: 'json'
    description: "File type to save the dataset as. Options are 'csv' and 'json'"
  deployment_validation:
    type: uri_file
    description: "Uri file containing information on if the Azure OpenAI deployments, if used, have been validated"
    optional: True

outputs:
  output_data:
    type: uri_folder
    description: "csv or jsonl file containing the question, answer, context, and metadata sets"

environment: azureml:llm-rag-embeddings@latest
code: '../src'  # nothing used from here

command: >-
  python -m azureml.rag.tasks.generate_qa
  --input-data '${{inputs.input_data}}'
  --output-data ${{outputs.output_data}}
  --dataset_size ${{inputs.dataset_size}}
  --chunk_batch_size ${{inputs.chunk_batch_size}}
  --llm_config '${{inputs.llm_config}}'
  --llm_connection '${{inputs.llm_connection}}'
  --output_format '${{inputs.output_format}}'
  --openai_api_version '${{inputs.openai_api_version}}'
  --openai_api_type '${{inputs.openai_api_type}}'