Class RepetitionPenaltyConfig

Namespace

Uralstech.UAI.LiteRTLM

public sealed class RepetitionPenaltyConfig : LiteRTLMNativeHandle

Inheritance

object

LiteRTLMNativeHandle

RepetitionPenaltyConfig

Inherited Members

LiteRTLMNativeHandle.Dispose()

Constructors

RepetitionPenaltyConfig()

Creates a LiteRT LM Repetition Penalty Config with default values (repetition_penalty = 1.0f, presence_penalty = 0.0f, frequency_penalty = 0.0f, window_size = 0, which means all history with no penalties active).

public RepetitionPenaltyConfig()

Remarks

When multiple penalties are configured and active, the order of application to output logits during decoding is:

1. Multiplicative penalty (repetition_penalty)
2. Subtractive penalties (presence_penalty and frequency_penalty).

The caller is responsible for disposing the wrapper using Dispose().

Exceptions

InvalidOperationException

Thrown if the native object could not be created.

Methods

ReleaseUnmanagedResources()

protected override void ReleaseUnmanagedResources()

SetFrequencyPenalty(float)

Sets the subtractive frequency penalty for the repetition penalty config.

public void SetFrequencyPenalty(float frequencyPenalty)

Parameters

frequencyPenalty float

A scalar subtracted from a token's logit, scaled linearly by the number of times that token has previously appeared inside the generated window history. Positive values discourage repetition, while negative values reward repeating tokens (OpenAI style). Defaults to 0.0f.

SetPresencePenalty(float)

Sets the subtractive presence penalty for the repetition penalty config.

public void SetPresencePenalty(float presencePenalty)

Parameters

presencePenalty float

A scalar subtracted from a token's logit if that token has appeared at least once inside the generated window history. Positive values discourage repetition, while negative values reward repeating tokens (OpenAI style). Defaults to 0.0f.

SetRepetitionPenalty(float)

Sets the multiplicative repetition penalty for the repetition penalty config.

public void SetRepetitionPenalty(float repetitionPenalty)

Parameters

repetitionPenalty float

A multiplicative penalty applied to a token's logit if that token has appeared at least once inside the generated window history (e.g., 1.0 = no penalty, 1.2 = moderate penalty). Positive logits are divided by this parameter, and negative logits are multiplied (HuggingFace style). The parameter must be >= 1.0f; values less than 1.0f are automatically clamped to 1.0f during execution.

SetWindowSize(int)

Sets the window size for the repetition penalty config.

public void SetWindowSize(int windowSize)

Parameters

windowSize int

The maximum number of recent tokens in generation history to consider when computing penalization. Tokens generated prior to this window are forgotten. A value of 0 means tracking all infinite generation history. Must be >= 0; negative values are clamped to 0 during execution.