Question Generators

Question generators turn seeds into questions. They run after seed generation and before labeling. Choose based on whether you need forward-looking questions, synthetic labels, or template-based generation.

QuestionGenerator

General-purpose question generation from seeds.

Parameter	Type	Required	Default	Description
instructions	str	No	—	Instructions for question generation
examples	list[str]	No	—	Example questions to guide generation
bad_examples	list[str]	No	—	Examples of questions to avoid
filter_	FilterCriteria or list	No	—	Optional filter criteria after generation
questions_per_seed	int	No	1	Questions to generate per seed
include_default_filter	bool	No	False	Include default filter for generated questions
answer_type	AnswerType	No	—	Expected answer type (Binary, Continuous, etc.)

ForwardLookingQuestionGenerator

Specialized for forward-looking forecasting questions. Same parameters as QuestionGenerator. Use when you want questions with explicit resolution dates and criteria.

ForwardLookingQuestionGenerator(
    instructions="Generate binary forecasting questions about Trump's actions and decisions.",
    examples=[
        "Will Trump impose 25% tariffs on all goods from Canada by February 1, 2025?",
        "Will Pete Hegseth be confirmed as Secretary of Defense by February 15, 2025?",
    ],
    answer_type=BinaryAnswerType(),
)

QuestionAndLabelGenerator

Generates both questions and labels in one step. Use when you don't need web search for labeling (e.g. synthetic labels from the model). No separate labeler is needed when using this generator.

Same parameters as QuestionGenerator. Omit labeler in your pipeline:

pipeline = QuestionPipeline(
    seed_generator=NewsSeedGenerator(...),
    question_generator=QuestionAndLabelGenerator(
        instructions="Generate binary forecasting questions with synthetic labels.",
        answer_type=BinaryAnswerType(),
    ),
    labeler=None,
)

TemplateQuestionGenerator

Generates questions from a template string. Use when you have structured data and a fixed question format.

Parameter	Type	Required	Default	Description
question_template	str	Yes	—	Template with placeholders. Supports {seed_text}, {seed}, {meta.label}

TemplateQuestionGenerator(
    question_template="Will {meta.label} be acquired by {meta.date_close}?",
)

Placeholders: {seed_text}, {seed}, and any key in meta (e.g. {meta.label}).

When to Use Each

Generator	Use case
ForwardLookingQuestionGenerator	Forecasting questions with resolution dates and criteria; most common for prediction tasks
QuestionGenerator	General questions from seeds
QuestionAndLabelGenerator	Questions + synthetic labels in one step; no web search labeling
TemplateQuestionGenerator	Structured questions from templates with {seed_text}, {meta.label}, etc.