koog-agents-optimization

0.1.0-SNAPSHOT

koog-agents-optimization/ai.koog.agents.optimization.optimizers.mipro/InstructionProposerConfig

InstructionProposerConfig

data class InstructionProposerConfig(val useDatasetSummary: Boolean = true, val programAware: Boolean = true, val useTaskDemos: Boolean = true, val numDemosInContext: Int = 3, val useTip: Boolean = true, val setTipRandomly: Boolean = true, val useInstructHistory: Boolean = false, val setHistoryRandomly: Boolean = false)

Configuration for InstructionProposer, toggling which context the meta-LLM receives.

Constructors

InstructionProposerConfig

constructor(useDatasetSummary: Boolean = true, programAware: Boolean = true, useTaskDemos: Boolean = true, numDemosInContext: Int = 3, useTip: Boolean = true, setTipRandomly: Boolean = true, useInstructHistory: Boolean = false, setHistoryRandomly: Boolean = false)

Properties

numDemosInContext

val numDemosInContext: Int = 3

maximum number of demonstrations to include when useTaskDemos is on.

val programAware: Boolean = true

include the agent's program/strategy description and code in the proposal context.

setHistoryRandomly

val setHistoryRandomly: Boolean = false

decide per run, at random, whether to include instruction history.

val setTipRandomly: Boolean = true

pick the tip at random rather than using the neutral default.

useDatasetSummary

val useDatasetSummary: Boolean = true

include an LLM-generated summary of the training dataset in the proposal context.

useInstructHistory

val useInstructHistory: Boolean = false

include previously proposed instructions and their scores as context.

val useTaskDemos: Boolean = true

include few-shot task demonstrations in the proposal context.

val useTip: Boolean = true

include a stylistic tip (from TIPS) to diversify generated instructions.