Struct EtaSampler 

pub struct EtaSampler {
    pub epsilon: f32,
    pub delta: f32,
}

Eta sampling — adaptively selects a probability cutoff based on distribution entropy.

The cutoff is max(epsilon, sqrt(exp(-H(p))) * delta) where H is the entropy. Tokens below the cutoff are discarded.

Fields

epsilon: f32

Minimum token probability (floor). Default: 0.0009.

delta: f32

Entropy scaling factor for adaptive threshold. Default: 0.07.

Implementations

impl EtaSampler

pub fn new(epsilon: f32, delta: f32) -> Self

Create a new eta sampler.

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

Sample a token index from raw logits.

Trait Implementations

impl Clone for EtaSampler

fn clone(&self) -> EtaSampler

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for EtaSampler

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

Formats the value using the given formatter.