Struct FineTuner 

pub struct FineTuner { /* private fields */ }

Embedding fine-tuner using contrastive learning losses.

Implementations

impl FineTuner

pub fn new(config: FinetuneConfig) -> Self

Create a fine-tuner with the given configuration.

pub fn compute_triplet_loss( &self, anchor: &[f32], positive: &[f32], negative: &[f32], ) -> f32

Compute triplet margin loss.

loss = max(0, d(a,p) - d(a,n) + margin)

pub fn compute_contrastive_loss(&self, a: &[f32], b: &[f32], label: f32) -> f32

Compute contrastive loss for a pair.

• label = 1.0 (similar): loss = d²
• label = 0.0 (dissimilar): loss = max(0, margin − d)²

pub fn compute_cosine_loss(&self, a: &[f32], b: &[f32], target: f32) -> f32

Compute cosine similarity loss: MSE between cosine_sim(a,b) and target.

pub fn step(&mut self, pairs: &[EmbeddingPair]) -> f32

Simulate one gradient step over the given pairs and return the mean loss.

pub fn train(&mut self, pairs: &[EmbeddingPair]) -> Vec<f32>

Run full training for config.epochs epochs and return the mean loss per epoch.

pub fn training_history(&self) -> &[TrainingStep]

Access the full training history.

pub fn total_steps(&self) -> usize

Total number of gradient steps recorded.