class

Llama::Sampler::Mirostat

Inherits Llama::Sampler::Base / Reference / Object

Mirostat sampler (version 1)

The Mirostat sampler dynamically adjusts the temperature to maintain a target entropy level in the generated text. This helps to produce consistent quality output regardless of the context.

Based on the paper: https://arxiv.org/abs/2007.14966

Example:

sampler = Llama::Sampler::Mirostat.new(32000, 42, 5.0, 0.1, 100)

Constructors

new(n_vocab : Int32, seed : UInt32, tau : Float32, eta : Float32, m : Int32)

Creates a new Mirostat sampler

Parameters:

  • n_vocab: Vocabulary size
  • seed: Random seed
  • tau: Target entropy (5.0 - 8.0 is a good range)
  • eta: Learning rate (0.1 is a good default)
  • m: Number of tokens for estimating entropy (100 is a good default)

Raises:

  • Llama::Error if the sampler cannot be created
Source