Is there documents or API reference about MIPROv2?

[Bingo-W]

I am a new user of dspy. But for some new API within dspy, such as the MIPROv2. I have not found any related documents. I am not sure if I miss them. And here are some specific questions:

• when is recommended to use MIPROv2?
• Can I use different datasets for searching for demonstration and evaluation for optimization, when using MIPROv2?

Thank you so much!

[stef41]

As of DSPy 2.5+, here's a comprehensive reference for MIPROv2:

What MIPROv2 does

MIPROv2 is a Bayesian prompt optimizer that uses Optuna-based search to find the best combination of:

1. Instructions — the natural language prompt for each predictor
2. Few-shot demonstrations — which examples to include in each predictor's prompt

It evaluates candidates against your metric and returns the best-performing program.

Basic usage

import dspy

optimizer = dspy.MIPROv2(metric=my_metric, auto="light")
optimized_program = optimizer.compile(my_program, trainset=trainset)

Key parameters

__init__ parameters

Parameter	Default	Description
metric	required	A callable metric(example, prediction) -> float that scores predictions
auto	"light"	Preset controlling search budget: "light", "medium", or "heavy". Controls num_candidates and num_trials automatically
num_candidates	None	Number of instruction/demo candidates to generate per predictor. Overrides auto
max_bootstrapped_demos	4	Max bootstrapped (model-generated) demos per predictor
max_labeled_demos	4	Max labeled (from trainset) demos per predictor
prompt_model	dspy.settings.lm	LM used to generate instruction candidates
task_model	dspy.settings.lm	LM used to evaluate candidates
num_threads	None	Parallelism for evaluation
seed	9	Random seed for reproducibility
verbose	False	Print optimization progress
metric_threshold	None	Early-stop if metric exceeds this

compile() parameters

Parameter	Default	Description
student	required	The DSPy program to optimize
trainset	required	List of dspy.Example for optimization
valset	None	Separate validation set (if None, splits from trainset)
teacher	None	A pre-optimized program to bootstrap demos from
num_trials	None	Number of Optuna trials (overrides auto)
minibatch	True	Use minibatch evaluation (faster)
minibatch_size	35	Size of evaluation minibatches
program_aware_proposer	True	Instruction proposer sees program structure
data_aware_proposer	True	Instruction proposer sees data samples

When to use MIPROv2

• Use MIPROv2 when you want to optimize both instructions and few-shot demos jointly. It's the most powerful optimizer in DSPy.
• Use auto="light" for quick iteration (fewer trials, ~minutes)
• Use auto="heavy" for production optimization (many trials, ~hours)
• For simpler cases where you only need good demos, BootstrapFewShot is faster.

To your specific question about different datasets

Can I use different datasets for searching for demonstration and evaluation for optimization?

Yes — use the valset parameter in compile():

optimized = optimizer.compile(
    my_program,
    trainset=demo_search_data,   # used for generating demo candidates
    valset=evaluation_data,       # used for evaluating candidates
)

If you don't pass valset, MIPROv2 automatically splits trainset into train/val portions.