Overview

Synthesizers transform the source samples read by a dataset generator’s data source into the output samples written to the generated dataset. Each synthesizer is defined inline in your dataset generator YAML under the synthesizers key.

# dataset_generators/my_generator.yaml
config_spec:
  - type: "model"
    key: "synthesizer_model"
    display_name: "Synthesizer Model
    default_value: "openai$gpt-4-1-nano
  - type: "dataset"
    key: "dataset_key"
    display_name: "Dataset"
    default_value: "countries"
definition:
  data_source:
    type: "dataset_samples"
    dataset_key: "<< config.dataset_key >>"
  synthesizers:
    - type: "llm"
      model_key: "<< config.synthesizer_model >>"
      use_structured_outputs: true
      user_prompt_template: >
        Write a trivia question about the capital of {{ source.country }}.
      sample_properties:
        question:
          type: string
        answer:
          type: string

A dataset generator can chain multiple synthesizers. They run in order, each consuming the samples produced by the previous one.

The synthesizers read their source samples from a data source, configured under the data_source key of the same dataset generator.

Available Synthesizers

Synthesizer	Type value	Description
Empty	empty	Produces no output samples. Used when the data source alone defines the dataset.
Drop Columns	drop_columns	Removes the named columns from each output sample, keeping all others.
Template	template	Renders output samples from a Jinja template with a grid of placeholder values.
Python	python	Generates output samples from each source sample using a custom Python synthesize function.
LLM	llm	Generates output samples from each source sample using a chat-completion model.
Question Answering	question_answering	Generates multiple-choice or open-ended QA samples from source text.
Claim Extraction	claim_extraction	Extracts atomic claims from a candidate answer using an LLM.
Claim Labeling	claim_labeling	Labels each extracted claim as required or optional for answering the question.
Link Claim Quotes	link_claim_quotes	Links each extracted claim to the verbatim quotes from the answer supporting it.