Struct DistillationConfig 

pub struct DistillationConfig {
    pub temperature: f32,
    pub alpha: f32,
    pub num_teachers: usize,
    pub student: StudentConfig,
    pub training: DistillTrainingConfig,
    pub output_dir: PathBuf,
}

Configuration for knowledge distillation training

Implements multi-teacher distillation via temperature-scaled KL divergence (Hinton et al. 2015). The loss function is:

L_distill = α * KL(softmax(z_s/T) || softmax(z_t/T)) + (1-α) * L_CE

Where:

• T is the temperature (higher = softer distributions)
• α is the balance between distillation and cross-entropy loss
• z_s and z_t are student and teacher logits

Fields

temperature: f32

Temperature for softmax (higher = softer probabilities) Typical values: 1.0-10.0. Default: 3.0

alpha: f32

Balance between distillation loss and CE loss α=1.0 means pure distillation, α=0.0 means pure CE Typical values: 0.5-0.9. Default: 0.7

num_teachers: usize

Number of teacher models for ensemble distillation

student: StudentConfig

Student model configuration

training: DistillTrainingConfig

Training hyperparameters

output_dir: PathBuf

Output directory for distilled model

Implementations

impl DistillationConfig

pub fn new() -> Self

Create a new distillation config

pub fn with_temperature(self, temperature: f32) -> Self

Builder: set temperature

pub fn with_alpha(self, alpha: f32) -> Self

Builder: set alpha (distillation weight)

pub fn with_teachers(self, num_teachers: usize) -> Self

Builder: set number of teachers

pub fn with_student(self, student: StudentConfig) -> Self

Builder: set student config

pub fn with_training(self, training: DistillTrainingConfig) -> Self

Builder: set training config

pub fn with_output_dir(self, output_dir: impl Into<PathBuf>) -> Self

Builder: set output directory

pub fn to_yaml(&self) -> String

Generate YAML configuration file for entrenar distillation

pub fn run_placeholder(&self) -> DistillationResult

Run placeholder distillation (simulates training)

Full distillation requires entrenar LLM feature and teacher model weights. This returns a placeholder result for testing the pipeline.

pub fn validate(&self) -> Result<(), String>

Validate configuration parameters

Errors

Returns error if parameters are invalid

Trait Implementations

impl Clone for DistillationConfig

fn clone(&self) -> DistillationConfig

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for DistillationConfig

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

Formats the value using the given formatter.

impl Default for DistillationConfig

fn default() -> Self

Returns the “default value” for a type.

impl<'de> Deserialize<'de> for DistillationConfig

fn deserialize<__D>(__deserializer: __D) -> Result<Self, __D::Error>

where __D: Deserializer<'de>,

Deserialize this value from the given Serde deserializer.

impl Serialize for DistillationConfig

fn serialize<__S>(&self, __serializer: __S) -> Result<__S::Ok, __S::Error>

where __S: Serializer,

Serialize this value into the given Serde serializer.