Struct MirostatV1Sampler 

pub struct MirostatV1Sampler {
    pub tau: f32,
    pub eta: f32,
    pub m: usize,
    /* private fields */
}

Mirostat v1 sampling — maintains target perplexity via feedback control.

Reference: Baktash et al., “Mirostat: A Neural Text Decoding Algorithm that Directly Controls Perplexity” (2020), https://arxiv.org/abs/2007.14966.

The algorithm:

1. Truncates the vocabulary to the top-m tokens.
2. Estimates the cross-entropy of the chosen token.
3. Updates mu (current estimate of target surprise) via eta.

Fields

tau: f32

Target surprise level (bits). Default: 5.0.

eta: f32

Learning rate for the feedback loop. Default: 0.1.

m: usize

Number of top candidates to consider. Typically vocab_size / 2.

Implementations

impl MirostatV1Sampler

pub fn new(tau: f32, eta: f32, m: usize) -> Self

Create a new v1 sampler.

pub fn sample(&mut self, logits: &[f32], rng: &mut LcgRng) -> usize

Sample a token index from raw logits, updating internal state.

pub fn reset(&mut self)

Reset the internal state to the initial value.

Trait Implementations

impl Clone for MirostatV1Sampler

fn clone(&self) -> MirostatV1Sampler

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for MirostatV1Sampler

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.