DSPy MIPROv2 Parameters Guide

This page enumerates and describes the MIPROv2 optimizer’s Constructor and Compiler parameters and behavior relate to each.

Background

This page exists to support the MIPROv2 walk through guide that describes MIPROv2’s step by step process and how to control those steps with these parameters.

This is unofficial documentation intended to aid in clarifying what each parameter in the class constructor and compiler do, so that you do not need to dig into the source code.

Based on DSPy Source Version: 3.0.0

For official MIPROv2 documentation see the optimizer page at DSPy.ai

Constructor Parameters

dspy.MIPROv2(
    metric: Callable,
    prompt_model: Any | None = None,
    task_model: Any | None = None,
    teacher_settings: dict | None = None,
    max_bootstrapped_demos: int = 4,
    max_labeled_demos: int = 4,
    auto: Literal["light", "medium", "heavy"] | None = "light",
    num_candidates: int | None = None,
    num_threads: int | None = None,
    max_errors: int | None = None,
    seed: int = 9,
    init_temperature: float = 0.5,
    verbose: bool = False,
    track_stats: bool = True,
    log_dir: str | None = None,
    metric_threshold: float | None = None,
)

Parameter	Type	Default	Description
metric	Callable	-	Function that compares LLM responses to expected answers from the validation set. Determines accuracy and scoring rules for optimization.
prompt_model	Any | None	None	LLM model for generating instruction variations (defaults to dspy.settings.lm).
task_model	Any | None	None	LLM model for running optimization trials (defaults to dspy.settings.lm).
teacher_settings	dict | None	{}	LLM model settings for running bootstrap demo evaluations.
max_bootstrapped_demos	int	4	Maximum number of examples selected from the training per demonstration set, that an LLM evaluates successfully.
max_labeled_demos	int	4	Maximum number of examples selected from the training set per demonstration set. These are randomly selected without LLM evaluation.
auto	Literal[‘light’, ‘medium’, ‘heavy’] | None	‘light’	Automatic optimization mode controlling intensity.
num_candidates	int | None	None	Number of demonstration sets and instruction candidates to create.
num_threads	int | None	None	Number of threads for parallel LLM calls during optimization.
max_errors	int | None	None	Maximum runtime errors allowed during evaluation.
seed	int	9	Random seed for reproducibility.
init_temperature	float	0.5	Sampling temperature for instruction generation.
verbose	bool	False	Enable verbose output.
track_stats	bool	True	Track optimization statistics.
log_dir	str | None	None	Directory for saving logs.
metric_threshold	float | None	None	Metric threshold for bootstrap demo selection.

compile Method Parameters

meth compile(
    student: Any,
    *,
    trainset: list,
    teacher: Any = None,
    valset: list | None = None,
    num_trials: int | None = None,
    max_bootstrapped_demos: int | None = None,
    max_labeled_demos: int | None = None,
    seed: int | None = None,
    minibatch: bool = True,
    minibatch_size: int = 35,
    minibatch_full_eval_steps: int = 5,
    program_aware_proposer: bool = True,
    data_aware_proposer: bool = True,
    view_data_batch_size: int = 10,
    tip_aware_proposer: bool = True,
    fewshot_aware_proposer: bool = True,
    requires_permission_to_run: bool = True, # deprecated
    provide_traceback: bool | None = None,
)

Parameter	Type	Default	Description
student	Any	-	DSPy program to optimize.
trainset	list	-	Training dataset used for selecting labeled demos and bootstrap demo examples.
teacher	Any	None	LLM model for running bootstrap demo evaluations. Usually a higher-quality model than the task_model.
valset	list | None	None	Validation dataset used for evaluating instruction and demo set combinations. Each trial runs the entire valset by default.
num_trials	int | None	None	Number of optimization trials to run. Required when not using auto mode. Each trial evaluates one specific instruction + demo set combination.
max_bootstrapped_demos	int | None	None	Override the constructor max_bootstrapped_demos value.
max_labeled_demos	int | None	None	Override the constructor max_labeled_demos value.
seed	int | None	None	Override random seed.
minibatch	bool	True	Enable minibatch evaluations to limit LLM runs per trial. Uses minibatch_size examples instead of full valset.
minibatch_size	int	35	Number of valset examples to use per trial when minibatch=True.
minibatch_full_eval_steps	int	5	Frequency of full evaluation trials on the most promising candidate.
program_aware_proposer	bool	True	Enable inclusion of the student program structure in instruction generation.
data_aware_proposer	bool	True	Enable LLM analysis of trainset characteristics for instruction generation. Uses view_data_batch_size.
view_data_batch_size	int	10	Number of training samples to analyze for data-aware instruction generation.
tip_aware_proposer	bool	True	Enable inclusion of randomly generated tips (e.g., “be creative”, “be concise”) in instruction generation.
fewshot_aware_proposer	bool	True	Enable inclusion of selected bootstrap demos in the instruction generation process. Not the same as bootstrapping in evaluation trials.
requires_permission_to_run	bool	True	Set Flase to disable MIPROv2 warnings about LLM run costs and required confirmation.
provide_traceback	bool | None	None	Provide tracebacks for errors.