Llama::Sampler::Penalties
Inherits Llama::Sampler::Base / Reference / Object
Penalties sampler
The Penalties sampler applies various penalties to token probabilities to reduce repetition and improve diversity in the generated text. It can penalize recently used tokens, frequent tokens, and more.
Example:
# Apply penalties to the last 64 tokens with a repetition penalty of 1.1
sampler = Llama::Sampler::Penalties.new(64, 1.1, 0.0, 0.0)Constructors
new(penalty_last_n : Int32, penalty_repeat : Float32, penalty_freq : Float32, penalty_present : Float32)
Creates a new Penalties sampler
Parameters:
- penalty_last_n: Last n tokens to penalize (0 = disable, -1 = context size)
- penalty_repeat: Repetition penalty (1.0 = disabled)
- penalty_freq: Frequency penalty (0.0 = disabled)
- penalty_present: Presence penalty (0.0 = disabled)
Raises:
- Llama::Error if the sampler cannot be created