Skip to main content
Sampling Strategies for budget forcing generation.

​
Classes

​
CLASS BudgetForcingSamplingStrategy

Sampling strategy that enforces a token budget for chain-of-thought reasoning. Extends RejectionSamplingStrategy with explicit control over the <think> block size and the answer block size. On each loop iteration, think_budget_forcing interleaves forced-thinking and final-answer generation, after which the standard rejection-sampling validation pass determines whether to accept or retry. Currently only supports the OllamaModelBackend. Args:
  • think_max_tokens: Tokens allocated for the thinking block. Defaults to 4096.
  • answer_max_tokens: Tokens allocated for the answer block. None means unbounded.
  • start_think_token: Token opening the thinking block. Defaults to "<think>".
  • end_think_token: Token closing the thinking block. Defaults to "</think>".
  • begin_response_token: Optional token opening the response block. Defaults to "".
  • end_response_token: Token closing the response block.
  • think_more_suffix: Suffix to force continued thinking. Empty string disables forcing.
  • answer_suffix: Suffix to elicit a final answer.
  • loop_budget: Rejection-sampling loop count. Must be > 0.
  • requirements: Requirements to validate against.
Methods:

​
FUNC sample 

sample(self, action: Component[S], context: Context, backend: Backend, requirements: list[Requirement] | None) -> SamplingResult[S]
This method performs a sampling operation based on the given instruction. Args:
  • action : The action object to be sampled.
  • context: The context to be passed to the sampling strategy.
  • backend: The backend used for generating samples.
  • requirements: List of requirements to test against (merged with global requirements).
  • validation_ctx: Optional context to use for validation. If None, validation_ctx = ctx.
  • format: output format for structured outputs.
  • model_options: model options to pass to the backend during generation / validation.
  • tool_calls: True if tool calls should be used during this sampling strategy.
  • show_progress: if true, a tqdm progress bar is used. Otherwise, messages will still be sent to flog.
Returns:
  • A result object indicating the success or failure of the sampling process.
Raises:
  • AssertionError: Asserts that all required components (repair, select_from_failure, validate, and generate) are provided before proceeding with the sampling.